Open-ended and AI-generating algorithms aim to continuously generate and solve increasingly complex tasks indefinitely, offering a promising path toward more general intelligence. To accomplish this grand vision, learning must occur within a vast array of potential tasks. Existing approaches to automatically generating environments are constrained within manually predefined, often narrow distributions of environments, limiting their ability to create any learning environment. To address this limitation, we introduce a novel framework, OMNI-EPIC, that augments previous work in Open-endedness via Models of human Notions of Interestingness (OMNI) with Environments Programmed in Code (EPIC). OMNI-EPIC leverages foundation models to autonomously generate code specifying the next learnable (i.e., not too easy or difficult for the agent’s current skill set) and interesting (e.g., worthwhile and novel) tasks. OMNI-EPIC generates both environments (e.g., an obstacle course) and reward functions (e.g., progress through the obstacle course quickly without touching red objects), enabling it, in principle, to create any simulatable learning task. We showcase the explosive creativity of OMNI-EPIC, which continuously innovates to suggest new, interesting learning challenges. We also highlight how OMNI-EPIC can adapt to reinforcement learning agents’ learning progress, generating tasks that are of suitable difficulty. Overall, OMNI-EPIC has the potential to endlessly create learnable and interesting environments, further propelling the development of self-improving AI systems and AI-Generating Algorithms.

Researchers are investing substantial effort in developing powerful general-purpose agents, wherein Foundation Models are used as modules within agentic systems (e.g. Chain-of-Thought, Self-Reflection, Toolformer). However, the history of machine learning teaches us that hand-designed solutions are eventually replaced by learned solutions. We describe a newly forming research area, Automated Design of Agentic Systems (ADAS), which aims to automatically create powerful agentic system designs, including inventing novel building blocks and/or combining them in new ways. We further demonstrate that there is an unexplored yet promising approach within ADAS where agents can be defined in code and new agents can be automatically discovered by a meta agent programming ever better ones in code. Given that programming languages are Turing Complete, this approach theoretically enables the learning of any possible agentic system: including novel prompts, tool use, workflows, and combinations thereof. We present a simple yet effective algorithm named Meta Agent Search to demonstrate this idea, where a meta agent iteratively programs interesting new agents based on an ever-growing archive of previous discoveries. Through extensive experiments across multiple domains including coding, science, and math, we show that our algorithm can progressively invent agents with novel designs that greatly outperform state-of-the-art hand-designed agents. Importantly, we consistently observe the surprising result that agents invented by Meta Agent Search maintain superior performance even when transferred across domains and models, demonstrating their robustness and generality. Provided we develop it safely, our work illustrates the potential of an exciting new research direction toward automatically designing ever-more powerful agentic systems to benefit humanity.

Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries---states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities. All our code is open-sourced at: https://github.com/conglu1997/intelligent-go-explore.

This work identifies a simple pre-training mechanism that leads to representations exhibiting better continual and transfer learning. This mechanism—the repeated resetting of weights in the last layer, which we nickname “zapping”—was originally designed for a meta-continual-learning procedure, yet we show it is surprisingly applicable in many settings beyond both meta-learning and continual learning. In our experiments, we wish to transfer a pre-trained image classifier to a new set of classes, in few shots. We show that our zapping procedure results in improved transfer accuracy and/or more rapid adaptation in both standard fine-tuning and continual learning settings, while being simple to implement and computationally efficient. In many cases, we achieve performance on par with state of the art meta-learning without needing the expensive higher-order gradients by using a combination of zapping and sequential learning. An intuitive explanation for the effectiveness of this zapping procedure is that representations trained with repeated zapping learn features that are capable of rapidly adapting to newly initialized classifiers. Such an approach may be considered a computationally cheaper type of, or alternative to, meta-learning rapidly adaptable features with higher-order gradients. This adds to recent work on the usefulness of resetting neural network parameters during training, and invites further investigation of this mechanism.

In many text-generation problems, users may prefer not only a single response, but a diverse range of high-quality outputs from which to choose. Quality-diversity (QD) search algorithms aim at such outcomes, by continually improving and diversifying a population of candidates. However, the applicability of QD to qualitative domains, like creative writing, has been limited by the difficulty of algorithmically specifying measures of quality and diversity. Interestingly, recent developments in language models (LMs) have enabled guiding search through \emph{AI feedback}, wherein LMs are prompted in natural language to evaluate qualitative aspects of text. Leveraging this development, we introduce Quality-Diversity through AI Feedback (QDAIF), wherein an evolutionary algorithm applies LMs to both generate variation and evaluate the quality and diversity of candidate text. When assessed on creative writing domains, QDAIF covers more of a specified search space with high-quality samples than do non-QD controls. Further, human evaluation of QDAIF-generated creative texts validates reasonable agreement between AI and human evaluation. Our results thus highlight the potential of AI feedback to guide open-ended search for creative and original solutions, providing a recipe that seemingly generalizes to many domains and modalities. In this way, QDAIF is a step towards AI systems that can independently search, diversify, evaluate, and improve, which are among the core skills underlying human society's capacity for innovation.

Open-ended algorithms aim to learn new, interesting behaviors forever. That requires a vast environment search space, but there are thus infinitely many possible tasks. Even after filtering for tasks the current agent can learn (i.e., learning progress), countless learnable yet uninteresting tasks remain (e.g., minor variations of previously learned tasks). An Achilles Heel of open-endedness research is the inability to quantify (and thus prioritize) tasks that are not just learnable, but also $\textit{interesting}$ (e.g., worthwhile and novel). We propose solving this problem by $\textit{Open-endedness via Models of human Notions of Interestingness}$ (OMNI). The insight is that we can utilize foundation models (FMs) as a model of interestingness (MoI), because they $\textit{already}$ internalize human concepts of interestingness from training on vast amounts of human-generated data, where humans naturally write about what they find interesting or boring. We show that FM-based MoIs improve open-ended learning by focusing on tasks that are both learnable $\textit{and interesting}$, outperforming baselines based on uniform task sampling or learning progress alone. This approach has the potential to dramatically advance the ability to intelligently select which tasks to focus on next (i.e., auto-curricula), and could be seen as AI selecting its own next task to learn, facilitating self-improving AI and AI-Generating Algorithms.

Reinforcement Learning from Human Feedback (RLHF) has shown potential in qualitative tasks where easily defined performance measures are lacking. However, there are drawbacks when RLHF is commonly used to optimize for average human preferences, especially in generative tasks that demand diverse model responses. Meanwhile, Quality Diversity (QD) algorithms excel at identifying diverse and high-quality solutions but often rely on manually crafted diversity metrics. This paper introduces Quality Diversity through Human Feedback (QDHF), a novel approach that progressively infers diversity metrics from human judgments of similarity among solutions, thereby enhancing the applicability and effectiveness of QD algorithms in complex and open-ended domains. Empirical studies show that QDHF significantly outperforms state-of-the-art methods in automatic diversity discovery and matches the efficacy of QD with manually crafted diversity metrics on standard benchmarks in robotics and reinforcement learning. Notably, in open-ended generative tasks, QDHF substantially enhances the diversity of text-to-image generation from a diffusion model and is more favorably received in user studies. We conclude by analyzing QDHF’s scalability, robustness, and quality of derived diversity metrics, emphasizing its strength in open-ended optimization tasks. Code and tutorials are available at https://liding.info/qdhf.

We introduce Genie, the first *generative interactive environment* trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual worlds described through text, synthetic images, photographs, and even sketches. At 11B parameters, Genie can be considered a *foundation world model*. It is comprised of a spatiotemporal video tokenizer, an autoregressive dynamics model, and a simple and scalable latent action model. Genie enables users to act in the generated environments on a frame-by-frame basis *despite training without any ground-truth action labels* or other domain specific requirements typically found in the world model literature. Further the resulting learned latent action space facilitates training agents to imitate behaviors from unseen videos, opening the path for training generalist agents of the future.

Standard reinforcement learning (RL) agents never intelligently explore like a human (i.e. taking into account complex domain priors and adapting quickly based on previous exploration). Across episodes, RL agents struggle to perform even simple exploration strategies, for example systematic search that avoids exploring the same location multiple times. This poor exploration limits performance on challenging domains. Meta-RL is a potential solution, as unlike standard RL, meta-RL can *learn* to explore, and potentially learn highly complex strategies far beyond those of standard RL, strategies such as experimenting in early episodes to learn new skills, or conducting experiments to learn about the current environment. Traditional meta-RL focuses on the problem of learning to optimally balance exploration and exploitation to maximize the *cumulative reward* of the episode sequence (e.g., aiming to maximize the total wins in a tournament -- while also improving as a player). We identify a new challenge with state-of-the-art cumulative-reward meta-RL methods. When optimal behavior requires exploration that sacrifices immediate reward to enable higher subsequent reward, existing state-of-the-art cumulative-reward meta-RL methods become stuck on the local optimum of failing to explore. Our method, First-Explore, overcomes this limitation by learning two policies: one to solely explore, and one to solely exploit. When exploring requires forgoing early-episode reward, First-Explore significantly outperforms existing cumulative meta-RL methods. By identifying and solving the previously unrecognized problem of forgoing reward in early episodes, First-Explore represents a significant step towards developing meta-RL algorithms capable of human-like exploration on a broader range of domains.

Language is often considered a key aspect of human thinking, providing us with exceptional abilities to generalize, explore, plan, replan, and adapt to new situations. However, Reinforcement Learning (RL) agents are far from human-level performance in any of these abilities. We hypothesize one reason for such cognitive deficiencies is that they lack the benefits of thinking in language and that we can improve AI agents by training them to $\textit{think like humans do}$. We introduce a novel Imitation Learning framework, Thought Cloning, where the idea is to not just clone the behaviors of human demonstrators, $\textit{but also the thoughts humans have as they perform these behaviors}$. While we expect Thought Cloning to truly shine at scale on internet-sized datasets (e.g. online videos with transcripts), here we conduct experiments in a domain where the thinking and action data are synthetically generated. Results reveal that Thought Cloning learns much faster than Behavioral Cloning and its performance advantage grows the further out of distribution test tasks are, highlighting its ability to better handle novel situations. Thought Cloning also provides important benefits for AI Safety and Interpretability, and makes it easier to debug and improve AI. Because we can observe the agent’s thoughts, we can (1) more easily diagnose why things are going wrong, making it easier to fix the problem, (2) steer the agent by correcting its thinking, or (3) prevent it from doing unsafe things it plans to do. Overall, by training agents $\textit{how to think}$ as well as behave, Thought Cloning creates safer, more powerful agents.

Pretraining on noisy, internet-scale datasets has been heavily studied as a technique for training models with broad, general capabilities for text, images, and other modalities. However, for many sequential decision domains such as robotics, video games, and computer use, publicly available data does not contain the labels required to train behavioral priors in the same way. We extend the internet-scale pretraining paradigm to sequential decision domains through semi-supervised imitation learning wherein agents learn to act by watching online unlabeled videos. Specifically, we show that with a small amount of labeled data we can train an inverse dynamics model accurate enough to label a huge unlabeled source of online data -- here, online videos of people playing Minecraft -- from which we can then train a general behavioral prior. Despite using the native human interface (mouse and keyboard at 20Hz), we show that this behavioral prior has nontrivial zero-shot capabilities and that it can be fine-tuned, with both imitation learning and reinforcement learning, to hard-exploration tasks that are impossible to learn from scratch via reinforcement learning. For many tasks our models exhibit human-level performance, and we are the first to report computer agents that can craft diamond tools, which can take proficient humans upwards of 20 minutes (24,000 environment actions) of gameplay to accomplish.

An important challenge in reinforcement learning is to solve multimodal problems, where agents have to act in qualitatively different ways depending on the circumstances. Because multimodal problems are often too difficult to solve directly, it is often helpful to define a curriculum, which is an ordered set of subtasks that can serve as the stepping stones for solving the overall problem. Unfortunately, choosing an effective ordering for these subtasks is difficult, and a poor ordering can reduce the performance of the learning process. Here, we provide a thorough introduction and investigation of the Combinatorial Multiobjective Evolutionary Algorithm (CMOEA), which allows all combinations of subtasks to be explored simultaneously. We compare CMOEA against three algorithms that can similarly optimize on multiple subtasks simultaneously: NSGA-II, NSGA-III, and ε-Lexicase Selection. The algorithms are tested on a function-optimization problem with two subtasks, a simulated multimodal robot locomotion problem with six subtasks, and a simulated robot maze-navigation problem where a hundred random mazes are treated as subtasks. On these problems, CMOEA either outperforms or is competitive with the controls. As a separate contribution, we show that adding a linear combination over all objectives can improve the ability of the control algorithms to solve these multimodal problems. Lastly, we show that CMOEA can leverage auxiliary objectives more effectively than the controls on the multimodal locomotion task. In general, our experiments suggest that CMOEA is a promising algorithm for solving multimodal problems.

Continual lifelong learning requires an agent or model to learn many sequentially ordered tasks, building on previous knowledge without catastrophically forgetting it. Much work has gone towards preventing the default tendency of machine learning models to catastrophically forget, yet virtually all such work involves manually-designed solutions to the problem. We instead advocate meta-learning a solution to catastrophic forgetting, allowing AI to learn to continually learn. Inspired by neuromodulatory processes in the brain, we propose A Neuromodulated Meta-Learning Algorithm (ANML). It differentiates through a sequential learning process to meta-learn an activation-gating function that enables context-dependent selective activation within a deep neural network. Specifically, a neuromodulatory (NM) neural network gates the forward pass of another (otherwise normal) neural network called the prediction learning network (PLN). The NM network also thus indirectly controls selective plasticity (i.e. the backward pass of) the PLN. ANML enables continual learning without catastrophic forgetting at scale: it produces state-of-the-art continual learning performance, sequentially learning as many as 600 classes (over 9,000 SGD updates).

Quality-Diversity (QD) algorithms, and MAP-Elites (ME) in particular, have proven very useful for a broad range of applications including enabling real robots to recover quickly from joint damage, solving strongly deceptive maze tasks or evolving robot morphologies to discover new gaits. However, present implementations of ME and other QD algorithms seem to be limited to low-dimensional controllers with far fewer parameters than modern deep neural network models. In this paper, we propose to leverage the efficiency of Evolution Strategies (ES) to scale MAP-Elites to high-dimensional controllers parameterized by large neural networks. We design and evaluate a new hybrid algorithm called MAP-Elites with Evolution Strategies (ME-ES) for post-damage recovery in a difficult high-dimensional control task where traditional ME fails. Additionally, we show that ME-ES performs efficient exploration, on par with state-of-the-art exploration algorithms in high-dimensional control tasks with strongly deceptive rewards.

Creating open-ended algorithms, which generate their own never-ending stream of novel and appropriately challenging learning opportunities, could help to automate and accelerate progress in machine learning. A recent step in this direction is the Paired Open-Ended Trailblazer (POET), an algorithm that generates and solves its own challenges, and allows solutions to goal-switch between challenges to avoid local optima. However, the original POET was unable to demonstrate its full creative potential because of limitations of the algorithm itself and because of external issues including a limited problem space and lack of a universal progress measure. Importantly, both limitations pose impediments not only for POET, but for the pursuit of open-endedness in general. Here we introduce and empirically validate two new innovations to the original algorithm, as well as two external innovations designed to help elucidate its full potential. Together, these four advances enable the most open-ended algorithmic demonstration to date. The algorithmic innovations are (1) a domain-general measure of how meaningfully novel new challenges are, enabling the system to potentially create and solve interesting challenges endlessly, and (2) an efficient heuristic for determining when agents should goal-switch from one problem to another (helping open-ended search better scale). Outside the algorithm itself, to enable a more definitive demonstration of open-endedness, we introduce (3) a novel, more flexible way to encode environmental challenges, and (4) a generic measure of the extent to which a system continues to exhibit open-ended innovation. Enhanced POET produces a diverse range of sophisticated behaviors that solve a wide range of environmental challenges, many of which cannot be solved through other means.

This paper investigates the intriguing question of whether we can create learning algorithms that automatically generate training data, learning environments, and curricula in order to help AI agents rapidly learn. We show that such algorithms are possible via Generative Teaching Networks (GTNs), a general approach that is, in theory, applicable to supervised, unsupervised, and reinforcement learning, although our experiments only focus on the supervised case. GTNs are deep neural networks that generate data and/or training environments that a learner (e.g. a freshly initialized neural network) trains on for a few SGD steps before being tested on a target task. We then differentiate \emph{through the entire learning process} via meta-gradients to update the GTN parameters to improve performance on the target task. This paper introduces GTNs, discusses their potential, and showcases that they can substantially accelerate learning. We also demonstrate a practical and exciting application of GTNs: accelerating the evaluation of candidate architectures for neural architecture search (NAS). GTN-NAS improves the NAS state of the art, finding higher performing architectures when controlling for the search proposal mechanism. GTN-NAS also is competitive with the overall state of the art approaches, which achieve top performance while using orders of magnitude less computation than typical NAS methods. Speculating forward, GTNs may represent a first step toward the ambitious goal of algorithms that generate their own training data and, in doing so, open a variety of interesting new research questions and directions.

Evolution provides a creative fount of complex and subtle adaptations that often surprise the scientists who discover them. However, the creativity of evolution is not limited to the natural world: Artificial organisms evolving in computational environments have also elicited surprise and wonder from the researchers studying them. The process of evolution is an algorithmic process that transcends the substrate in which it occurs. Indeed, many researchers in the field of digital evolution can provide examples of how their evolving algorithms and organisms have creatively subverted their expectations or intentions, exposed unrecognized bugs in their code, produced unexpectedly adaptations, or engaged in behaviors and outcomes, uncannily convergent with ones found in nature. Such stories routinely reveal surprise and creativity by evolution in these digital worlds, but they rarely fit into the standard scientific narrative. Instead they are often treated as mere obstacles to be overcome, rather than results that warrant study in their own right. Bugs are fixed, experiments are refocused, and one-off surprises are collapsed into a single data point. The stories themselves are traded among researchers through oral tradition, but that mode of information transmission is inefficient and prone to error and outright loss. Moreover, the fact that these stories tend to be shared only among practitioners means that many natural scientists do not realize how interesting and lifelike digital organisms are and how natural their evolution can be. To our knowledge, no collection of such anecdotes has been published before. This article is the crowd-sourced product of researchers in the fields of artificial life and evolutionary computation who have provided first-hand accounts of such cases. It thus serves as a written, fact-checked collection of scientifically important and even entertaining stories. In doing so we also present here substantial evidence that the existence and importance of evolutionary surprises extends beyond the natural world, and may indeed be a universal property of all complex evolving systems.

Designing evolutionary algorithms capable of uncovering highly evolvable representations is an open challenge in evolutionary computation; such evolvability is important in practice, because it accelerates evolution and enables fast adaptation to changing circumstances. This paper introduces evolvability ES, an evolutionary algorithm designed to explicitly and efficiently optimize for evolvability, i.e. the ability to further adapt. The insight is that it is possible to derive a novel objective in the spirit of natural evolution strategies that maximizes the diversity of behaviors exhibited when an individual is subject to random mutations, and that efficiently scales with computation. Experiments in 2-D and 3-D locomotion tasks highlight the potential of evolvability ES to generate solutions with tens of thousands of parameters that can quickly be adapted to solve different tasks and that can productively seed further evolution. We further highlight a connection between evolvability in EC and a recent and popular gradient-based meta-learning algorithm called MAML; results show that evolvability ES can perform competitively with MAML and that it discovers solutions with distinct properties. The conclusion is that evolvability ES opens up novel research directions for studying and exploiting the potential of evolvable representations for deep neural networks.

How can progress in machine learning and reinforcement learning be automated to generate its own never-ending curriculum of challenges without human intervention? The recent emergence of quality diversity (QD) algorithms offers a glimpse of the potential for such continual open-ended invention. For example, novelty search showcases the benefits of explicit novelty pressure, MAP-Elites and Innovation Engines highlight the advantage of explicit elitism within niches in an otherwise divergent process, and minimal criterion coevolution (MCC) reveals that problems and solutions can coevolve divergently. The Paired Open-Ended Trailblazer (POET) algorithm introduced in this paper combines these principles to produce a practical approach to generating an endless progression of diverse and increasingly challenging environments while at the same time explicitly optimizing their solutions. An intriguing implication is the opportunity to transfer solutions among environments, reflecting the view that innovation is a circuitous and unpredictable process. POET is tested in a 2-D obstacles course domain, where it generates diverse and sophisticated behaviors that create and solve a wide range of environmental challenges, many of which cannot be solved by direct optimization, or by a direct-path curriculum-building control algorithm. We hope that POET will inspire a new push towards open-ended discovery across many domains.

Deep learning provides a powerful tool for robotic perception in the open world. However, real-world robotic systems, especially mobile robots, must be able to react intelligently and safely even in unexpected circumstances. This requires a system that knows what it knows, and can estimate its own uncertainty for unfamiliar, out-of-distribution observations. Approximate Bayesian approaches are commonly used to estimate uncertainty for neural network predictions, but struggle with out-of-distribution observations. Generative models can in principle detect out-of-distribution observations as those with a low estimated density, but overly pessimistic as an uncertainty measure, since the mere presence of an out-of-distribution input does not by itself indicate an unsafe situation. Intuitively, we would like a perception system that can detect when task-salient parts of the image are unfamiliar or uncertain, while ignoring task-irrelevant features. In this paper, we present a method for uncertainty-aware robotic perception that combines generative modeling and model uncertainty. Our method estimates an uncertainty measure about the model's prediction, taking into account an explicit generative model of the observation distribution to handle out-of-distribution inputs. We evaluate our method on an action-conditioned collision prediction task with both simulated and real data, and demonstrate that our approach improves on a variety of Bayesian neural network techniques.

Much human and computational effort has aimed to improve how deep reinforcement learning (DRL) algorithms perform on benchmarks such as the Atari Learning Environment. Comparatively less effort has focused on understanding what has been learned by such methods, and investigating and comparing the representations learned by different families of DRL algorithms. Sources of friction include the onerous computational requirements, and general logistical and architectural complications for running DRL algorithms at scale. We lessen this friction, by (1) training several algorithms at scale and releasing trained models, (2) integrating with a previous DRL model release, and (3) releasing code that makes it easy for anyone to load, visualize, and analyze such models. This paper introduces the Atari Zoo framework, which contains models trained across benchmark Atari games, in an easy-to-use format, as well as code that implements common modes of analysis and connects such models to a popular neural network visualization library. Further, to demonstrate the potential of this dataset and software package, we show initial quantitative and qualitative comparisons between the performance and representations of several DRL algorithms, highlighting interesting and previously unknown distinctions between them.

While neuroevolution (evolving neural networks) has been successful across a variety of domains from reinforcement learning, to artificial life, to evolutionary robotics, it is rarely applied to large, deep neural networks. A central reason is that while random mutation generally works in low dimensions, a random perturbation of thousands or millions of weights will likely break existing functionality. This paper proposes a solution: a family of safe mutation (SM) operators that facilitate exploration without dramatically altering network behavior or requiring additional interaction with the environment. The most effective SM variant scales the degree of mutation of each individual weight according to the sensitivity of the network's outputs to that weight, which requires computing the gradient of outputs with respect to the weights (instead of the gradient of error, as in conventional deep learning). This safe mutation through gradients (SM-G) operator dramatically increases the ability of a simple genetic algorithm-based neuroevolution method to find solutions in high-dimensional domains that require deep and/or recurrent neural networks, including domains that require processing raw pixels. By improving our ability to evolve deep neural networks, this new safer approach to mutation expands the scope of domains amenable to neuroevolution.

An evolution strategy (ES) variant based on a simplification of a natural evolution strategy recently attracted attention because it performs surprisingly well in challenging deep reinforcement learning domains. It searches for neural network parameters by generating perturbations to the current set of parameters, checking their performance, and moving in the aggregate direction of higher reward. Because it resembles a traditional finite-difference approximation of the reward gradient, it can naturally be confused with one. However, this ES optimizes for a different gradient than just reward: It optimizes for the average reward of the entire population, thereby seeking parameters that are robust to perturbation. This difference can channel ES into distinct areas of the search space relative to gradient descent, and also consequently to networks with distinct properties. This unique robustness-seeking property, and its consequences for optimization, are demonstrated in several domains. They include humanoid locomotion, where networks from policy gradient-based reinforcement learning are significantly less robust to parameter perturbation than ES-based policies solving the same task. While the implications of such robustness and robustness-seeking remain open to further study this work's main contribution is to highlight such differences and their potential importance.

How can we build agents that keep learning from experience, quickly and efficiently, after their initial training? Here we take inspiration from the main mechanism of learning in biological brains: synaptic plasticity, carefully tuned by evolution to produce efficient lifelong learning. We show that plasticity, just like connection weights, can be optimized by gradient descent in large (millions of parameters) recurrent networks with Hebbian plastic connections. First, recurrent plastic networks with more than two million parameters can be trained to memorize and reconstruct sets of novel, high-dimensional (1000+ pixels) natural images not seen during training. Crucially, traditional non-plastic recurrent networks fail to solve this task. Furthermore, trained plastic networks can also solve generic meta-learning tasks such as the Omniglot task, with competitive results and little parameter overhead. Finally, in reinforcement learning settings, plastic networks outperform non-plastic equivalent in a maze exploration task. We conclude that differentiable plasticity may provide a powerful novel approach to the learning-to-learn problem.

Evolution strategies (ES) are a family of black-box optimization algorithms able to train deep neural networks roughly as well as Q-learning and policy gradient methods on challenging deep reinforcement learning (RL) problems, but are much faster (e.g. hours vs. days) because they parallelize better. However, many RL problems require directed exploration because they have reward functions that are sparse or deceptive (i.e. contain local optima), and it is unknown how to encourage such exploration with ES. Here we show that algorithms that have been invented to promote directed exploration in small-scale evolved neural networks via populations of exploring agents, specifically novelty search (NS) and quality diversity (QD) algorithms, can be hybridized with ES to improve its performance on sparse or deceptive deep RL tasks, while retaining scalability. Our experiments confirm that the resultant new algorithms, NS-ES and two QD algorithms, NSR-ES and NSRA-ES, avoid local optima encountered by ES to achieve higher performance on Atari and simulated robots learning to walk around a deceptive trap. This paper thus introduces a family of fast, scalable algorithms for reinforcement learning that are capable of directed exploration. It also adds this new family of exploration algorithms to the RL toolbox and raises the interesting possibility that analogous algorithms with multiple simultaneous paths of exploration might also combine well with existing RL algorithms outside ES.

Many believe that an essential component for the discovery of the tremendous diversity in natural organisms was the evolution of evolvability, whereby evolution speeds up its ability to innovate by generating a more adaptive pool of offspring. One hypothesized mechanism for evolvability is developmental canalization, wherein certain dimensions of variation become more likely to be traversed and others are prevented from being explored (e.g., offspring tend to have similar-size legs, and mutations affect the length of both legs, not each leg individually). While ubiquitous in nature, canalization is rarely reported in computational simulations of evolution, which deprives us of in silico examples of canalization to study and raises the question of which conditions give rise to this form of evolvability. Answering this question would shed light on why such evolvability emerged naturally, and it could accelerate engineering efforts to harness evolution to solve important engineering challenges. In this article, we reveal a unique system in which canalization did emerge in computational evolution. We document that genomes entrench certain dimensions of variation that were frequently explored during their evolutionary history. The genetic representation of these organisms also evolved to be more modular and hierarchical than expected by chance, and we show that these organizational properties correlate with increased fitness. Interestingly, the type of computational evolutionary experiment that produced this evolvability was very different from traditional digital evolution in that there was no objective, suggesting that open-ended, divergent evolutionary processes may be necessary for the evolution of evolvability.

Generating high-resolution, photo-realistic images has been a long-standing goal in machine learning. Recently, Nguyen et al. [37] showed one interesting way to synthesize novel images by performing gradient ascent in the latent space of a generator network to maximize the activations of one or multiple neurons in a separate classifier network. In this paper we extend this method by introducing an additional prior on the latent code, improving both sample quality and sample diversity, leading to a state-of-the-art generative model that produces high quality images at higher resolutions (227 × 227) than previous generative models, and does so for all 1000 ImageNet categories. In addition, we provide a unified probabilistic interpretation of related activation maximization methods and call the general class of models Plug and Play Generative Networks. PPGNs are composed of 1) a generator network G that is capable of drawing a wide range of image types and 2) a replaceable condition network C that tells the generator what to draw. We demonstrate the generation of images conditioned on a class (when C is an ImageNet or MIT Places classification network) and also conditioned on a caption (when C is an image captioning network). Our method also improves the state of the art of Multifaceted Feature Visualization [40], which generates the set of synthetic inputs that activate a neuron in order to better understand how deep neural networks operate. Finally, we show that our model performs reasonably well at the task of image inpainting. While image models are used in this paper, the approach is modality-agnostic and can be applied to many types of data.

One of the most intriguing questions in evolution is how organisms exhibit suitable phenotypic variation to rapidly adapt in novel selective environments. Such variability is crucial for evolvability, but poorly understood. In particular, how can natural selection favour developmental organisations that facilitate adaptive evolution in previously unseen environments? Such a capacity suggests foresight that is incompatible with the short-sighted concept of natural selection. A potential resolution is provided by the idea that evolution may discover and exploit information not only about the particular phenotypes selected in the past, but their underlying structural regularities: new phenotypes, with the same underlying regularities, but novel particulars, may then be useful in new environments. If true, we still need to understand the conditions in which natural selection will discover such deep regularities rather than exploiting 'quick fixes' (i.e., fixes that provide adaptive phenotypes in the short term, but limit future evolvability). Here we argue that the ability of evolution to discover such regularities is formally analogous to learning principles, familiar in humans and machines, that enable generalisation from past experience. Conversely, natural selection that fails to enhance evolvability is directly analogous to the learning problem of over-fitting and the subsequent failure to generalise. We support the conclusion that evolving systems and learning systems are different instantiations of the same algorithmic principles by showing that existing results from the learning domain can be transferred to the evolution domain. Specifically, we show that conditions that alleviate over-fitting in learning systems successfully predict which biological conditions (e.g., environmental variation, regularity, noise or a pressure for developmental simplicity) enhance evolvability. This equivalence provides access to a well-developed theoretical framework from learning theory that enables a characterisation of the general conditions for the evolution of evolvability.

Many argue that to evolve artificial intelligence that rivals that of natural animals, we need to evolve neural networks that are structurally organized in that they exhibit modularity, regularity, and hierarchy. It was recently shown that a cost for network connections, which encourages the evolution of modularity, can be combined with an indirect encoding, which encourages the evolution of regularity, to evolve networks that are both modular and regular. However, the bias towards regularity from indirect encodings may prevent evolution from independently optimizing different modules to perform different functions, unless modularity in the phenotype is aligned with modularity in the genotype. We test this hypothesis on two multi-modal problems---a pattern recognition task and a robotics task---that each require different phenotypic modules. In general, we find that performance is improved only when genotypic and phenotypic modularity are encouraged simultaneously, though the role of alignment remains unclear. In addition, intuitive manual decompositions fail to provide the performance benefits of automatic methods on the more challenging robotics problem, emphasizing the importance of automatic, rather than manual, decomposition methods. These results suggest encouraging modularity in both the genotype and phenotype as an important step towards solving large-scale multi-modal problems, but also indicate that more research is required before we can evolve structurally organized networks to solve tasks that require multiple, different neural modules.

One hallmark of natural organisms is their significant evolvability, i.e.,their increased potential for further evolution. However, reproducing such evolvability in artificial evolution remains a challenge, which both reduces the performance of evolutionary algorithms and inhibits the study of evolvable digital phenotypes. Although some types of selection in evolutionary computation indirectly encourage evolvability, one unexplored possibility is to directly select for evolvability. To do so, we estimate an individual's future potential for diversity by calculating the behavioral diversity of its immediate offspring, and select organisms with increased offspring variation. While the technique is computationally expensive, we hypothesized that direct selection would better encourage evolvability than indirect methods. Experiments in two evolutionary robotics domains confirm this hypothesis: in both domains, such Evolvability Search produces solutions with higher evolvability than those produced with Novelty Search or traditional objective-based search algorithms. Further experiments demonstrate that the higher evolvability produced by Evolvability Search in a training environment also generalizes, producing higher evolvability in a new test environment without further selection. Overall, Evolvability Search enables generating evolvability more easily and directly, facilitating its study and understanding, and may inspire future practical algorithms that increase evolvability without significant computational overhead.

Many animals predict the outcomes of their actions by internal models. Such "forward models'' enable animals to rapidly simulate many actions without performing them to choose an appropriate action. Robots would similarly benefit from forward models. However, such models must change over time to account for changes in the environment or body, such as injury. Thus, forward models must not only be accurate, but also adaptable. Neural networks can learn complex functions with high accuracy, hence they are suitable candidates to build forward models for robots. Generally, neural networks are static, which means once they pass the training phase, their weights remain unchanged and they thus cannot adapt themselves if something about the world or their body changes. Plastic neural networks change their connections over time via local learning rules (e.g. Hebbian rule) and can thus deal with unforeseen changes. A more complex, yet still biologically-inspired, technique is neuromodulation, which can change per-connection learning rates in different contexts. In this paper, we test the hypothesis that neuromodulation may improve the evolution of forward models because it can heighten learning after drastic changes such as injury. We compare forward models evolved with neuromodulation to those evolved with static neural networks and Hebbian plastic neural networks. The results show that neuromodulation produces forward models that can adapt to changes significantly better than the controls. Our findings suggest that neuromodulation is an effective tool for enabling robots (and artificial intelligence agents, more generally) to have more adaptable, effective forward models.

The recently introduced Intelligent Trial and Error algorithm (IT&E) both improves the ability to automatically generate controllers that transfer to real robots, and enables robots to creatively adapt to damage in less than 2 minutes. A key component of IT&E is a new evolutionary algorithm called MAP-Elites, which creates a behavior-performance map that is provided as a set of "creative" ideas to an online learning algorithm. To date, all experiments with MAP-Elites have been performed with a directly encoded list of parameters: it is therefore unknown how MAP-Elites would behave with more advanced encodings, like HyperNeat and SUPG. In addition, because we ultimately want robots that respond to their environments via sensors, we investigate the ability of MAP-Elites to evolve closed-loop controllers, which are more complicated, but also more powerful. Our results show that the encoding critically impacts the quality of the results of MAP-Elites, and that the differences are likely linked to the locality of the encoding (the likelihood of generating a similar behavior after a single mutation). Overall, these results improve our understanding of both the dynamics of the MAP-Elites algorithm and how to best harness MAP-Elites to evolve effective and adaptable robotic controllers.

As the power and capabilities of Artificial Neural Networks (ANNs) grow, so do their size and complexity. To both decipher and improve ANNs, we need to build better tools that help us understand their inner workings. To that end, we introduce an algorithm called Subsets Regression on network Connectivity (SRC). SRC allows us to prune away unimportant nodes and connections in ANNs, revealing a core functional network (CFN) that is simpler and thus easier to analyze. SRC can also identify functional modules within an ANN. We demonstrate SRC's capabilities on both directly and indirectly encoded ANNs evolved to solve a modular problem. In many of the cases when evolution produces a highly entangled, non-modular ANN, SRC reveals that a CFN is hidden within these networks that is actually sparse and modular. That finding will substantially impact the sizable and ongoing research into the evolution of modularity and will encourage researchers to revisit previous results on that topic. We also show that the SRC algorithm can more accurately estimate the modularity Q-Score of a network than state-of-the-art approaches. Overall, SRC enables us to greatly simplify ANNs in order to better understand and improve them, and reveals that they often contain hidden modular structures within.

Deep neural networks (DNNs) have demonstrated state-of-the-art results on many pattern recognition tasks, especially vision classification problems. Understanding the inner workings of such computational brains is both fascinating basic science that is interesting in its own right---similar to why we study the human brain---and will enable researchers to further improve DNNs. One path to understanding how a neural network functions internally is to study what each of its neurons has learned to detect. One such method is called activation maximization, which synthesizes an input (e.g. an image) that highly activates a neuron. Here we dramatically improve the qualitative state of the art of activation maximization by harnessing a powerful, learned prior: a deep generator network. The algorithm (1) generates qualitatively state-of-the-art synthetic images that look almost real, (2) reveals the features learned by each neuron in an interpretable way, (3) generalizes well to new datasets and somewhat well to different network architectures without requiring the prior to be relearned, and (4) can be considered as a high-quality generative method (in this case, by generating novel, creative, interesting, recognizable images).

We present a survey of the first 21 years of web-based artificial life (WebAL) research and applications, broadly construed to include the many different ways in which artificial life and web technologies might intersect. Our survey covers the period from 1994-when the first WebAL work appeared-up to the present day, together with a brief discussion of relevant precursors. We examine recent projects, from 2010-2015, in greater detail in order to highlight the current state of the art. We follow the survey with a discussion of common themes and methodologies that can be observed in recent work and identify a number of likely directions for future work in this exciting area.

The Achilles Heel of stochastic optimization algorithms is getting trapped on local optima. Novelty Search mitigates this problem by encouraging exploration in all interesting directions by replacing the performance objective with a reward for novel behaviors. This reward for novel behaviors has traditionally required a human-crafted, behavioral distance function. While Novelty Search is a major conceptual breakthrough and outperforms traditional stochastic optimization on certain problems, it is not clear how to apply it to challenging, high-dimensional problems where specifying a useful behavioral distance function is difficult. For example, in the space of images, how do you encourage novelty to produce hawks and heroes instead of endless pixel static? Here we propose a new algorithm, the Innovation Engine, that builds on Novelty Search by replacing the human-crafted behavioral distance with a Deep Neural Network (DNN) that can recognize interesting differences between phenotypes. The key insight is that DNNs can recognize similarities and differences between phenotypes at an abstract level, wherein novelty means interesting novelty. For example, a DNN-based novelty search in the image space does not explore in the low-level pixel space, but instead creates a pressure to create new types of images (e.g., churches, mosques, obelisks, etc.). Here, we describe the long-term vision for the Innovation Engine algorithm, which involves many technical challenges that remain to be solved. We then implement a simplified version of the algorithm that enables us to explore some of the algorithm's key motivations. Our initial results, in the domain of images, suggest that Innovation Engines could ultimately automate the production of endless streams of interesting solutions in any domain: for example, producing intelligent software, robot controllers, optimized physical components, and art.

Hierarchical organization-the recursive composition of sub-modules-is ubiquitous in biological networks, including neural, metabolic, ecological, and genetic regulatory networks, and in human-made systems, such as large organizations and the Internet. To date, most research on hierarchy in networks has been limited to quantifying this property. However, an open, important question in evolutionary biology is why hierarchical organization evolves in the first place. It has recently been shown that modularity evolves because of the presence of a cost for network connections. Here we investigate whether such connection costs also tend to cause a hierarchical organization of such modules. In computational simulations, we find that networks without a connection cost do not evolve to be hierarchical, even when the task has a hierarchical structure. However, with a connection cost, networks evolve to be both modular and hierarchical, and these networks exhibit higher overall performance and evolvability (i.e. faster adaptation to new environments). Additional analyses confirm that hierarchy independently improves adaptability after controlling for modularity. Overall, our results suggest that the same force-the cost of connections-promotes the evolution of both hierarchy and modularity, and that these properties are important drivers of network performance and adaptability. In addition to shedding light on the emergence of hierarchy across the many domains in which it appears, these findings will also accelerate future research into evolving more complex, intelligent computational brains in the fields of artificial intelligence and robotics.

Deep neural networks (DNNs) have recently been achieving state-of-the-art performance on a variety of pattern-recognition tasks, most notably visual classification problems. Given that DNNs are now able to classify objects in images with near-human-level performance, questions naturally arise as to what differences remain between computer and human vision. A recent study [30] revealed that changing an image (e.g. of a lion) in a way imperceptible to humans can cause a DNN to label the image as something else entirely (e.g. mislabeling a lion a library). Here we show a related result: it is easy to produce images that are completely unrecognizable to humans, but that state-of-the-art DNNs believe to be recognizable objects with 99.99% confidence (e.g. labeling with certainty that white noise static is a lion). Specifically, we take convolutional neural networks trained to perform well on either the ImageNet or MNIST datasets and then find images with evolutionary algorithms or gradient ascent that DNNs label with high confidence as belonging to each dataset class. It is possible to produce images totally unrecognizable to human eyes that DNNs believe with near certainty are familiar objects, which we call “fooling images” (more generally, fooling examples). Our results shed light on interesting differences between human vision and current DNNs, and raise questions about the generality of DNN computer vision.

The Achilles Heel of stochastic optimization algorithms is getting trapped on local optima. Novelty Search avoids this problem by encouraging a search in all interesting directions. That occurs by replacing a performance objective with a reward for novel behaviors, as defined by a human-crafted, and often simple, behavioral distance function. While Novelty Search is a major conceptual breakthrough and outperforms traditional stochastic optimization on certain problems, it is not clear how to apply it to challenging, high-dimensional problems where specifying a useful behavioral distance function is difficult. For example, in the space of images, how do you encourage novelty to produce hawks and heroes instead of endless pixel static? Here we propose a new algorithm, the Innovation Engine, that builds on Novelty Search by replacing the human-crafted behavioral distance with a Deep Neural Network (DNN) that can recognize interesting differences between phenotypes. The key insight is that DNNs can recognize similarities and differences between phenotypes at an abstract level, wherein novelty means interesting novelty. For example, a novelty pressure in image space does not explore in the low-level pixel space, but instead creates a pressure to create new types of images (e.g. churches, mosques, obelisks, etc.). Here we describe the long-term vision for the Innovation Engine algorithm, which involves many technical challenges that remain to be solved. We then implement a simplified version of the algorithm that enables us to explore some of the algorithm's key motivations. Our initial results, in the domain of images, suggest that Innovation Engines could ultimately automate the production of endless streams of interesting solutions in any domain: e.g. producing intelligent software, robot controllers, optimized physical components, and art.

A long-standing goal in artificial intelligence is creating agents that can learn a variety of different skills for different problems. In the artificial intelligence subfield of neural networks, a barrier to that goal is that when agents learn a new skill they typically do so by losing previously acquired skills, a problem called catastrophic forgetting. That occurs because, to learn the new task, neural learning algorithms change connections that encode previously acquired skills. How networks are organized critically affects their learning dynamics. In this paper, we test whether catastrophic forgetting can be reduced by evolving modular neural networks. Modularity intuitively should reduce learning interference between tasks by separating functionality into physically distinct modules in which learning can be selectively turned on or off. Modularity can further improve learning by having a reinforcement learning module separate from sensory processing modules, allowing learning to happen only in response to a positive or negative reward. In this paper, learning takes place via neuromodulation, which allows agents to selectively change the rate of learning for each neural connection based on environmental stimuli (e.g. to alter learning in specific locations based on the task at hand). To produce modularity, we evolve neural networks with a cost for neural connections. We show that this connection cost technique causes modularity, confirming a previous result, and that such sparsely connected, modular networks have higher overall performance because they learn new skills faster while retaining old skills more and because they have a separate reinforcement learning module. Our results suggest (1) that encouraging modularity in neural networks may help us overcome the long-standing barrier of networks that cannot learn new skills without forgetting old ones, and (2) that one benefit of the modularity ubiquitous in the brains of natural animals might be to alleviate the problem of catastrophic forgetting.

The embodied cognition paradigm emphasizes that both bodies and brains combine to produce complex behaviors, in contrast to the traditional view that the only seat of intelligence is the brain. Despite recent excitement about embodied cognition, brains and bodies remain thought of, and implemented as, two separate entities that merely interface with one another to carry out their respective roles. Previous research co-evolving bodies and brains has simulated the physics of bodies that collect sensory information and pass that information on to disembodied neural networks, which then processes that information and return motor commands. Biological animals, in contrast, produce behavior through physically embedded control structures and a complex and continuous interplay between neural and mechanical forces. In addition to the electrical pulses flowing through the physical wiring of the nervous system, the heart elegantly combines control with actuation, as the physical properties of the tissue itself (or defects therein) determine the actuation of the organ. Inspired by these phenomena from cardiac electrophysiology (the study of the electrical properties of heart tissue), we introduce electrophysiological robots, whose behavior is dictated by electrical signals flowing though the tissue cells of soft robots. Here we describe these robots and how they are evolved. Videos and images of these robots reveal lifelike behaviors despite the added challenge of having physically embedded control structures. We also provide an initial experimental investigation into the impact of different implementation decisions, such as alternatives for sensing, actuation, and locations of central pattern generators. Overall, this paper provides a first step towards removing the chasm between bodies and brains to encourage further research into physically realistic embodied cognition. Introduction and Background The fields of evolutionary robotics and artificial life have seen a great deal of emphasis on embodied cognition in recent years [Cheney et al. (2013); Bongard (2013); Rieffel et al. (2013); Auerbach and Bongard (2012); Hiller and Lipson (2012a); Lehman and Stanley (2011); Auerbach and Bongard (2010a,b); Pfeifer et al. (2007); Hornby et al. (2001); Lipson and Pollack (2000)]. There is even a paradigm called embodied cognition, which argues that the specifics of the embodiment (such as the morphology) are Figure 1: Current flowing through an evolved creature. The legend for voltage within each cell (colors) is given in Fig. 3. vital parts of the resulting behavior of the system: It argues that the co-evolutionary connection between body and brain is more deeply intertwined than the body simply acting as a minimal physical interface between the brain and the environment [Pfeifer and Bongard (2006)]. Recent work in evolutionary robotics has shown that complex behaviors can arise when co-evolving bodies and brains. At one end of the spectrum, Auerbach and Bongard (2010b) demonstrated the evolution of physical structures that had no joints or actuators, and evolved to cover the largest distance in a controlled fall due to gravity. While that work exemplifies the evolution of behavior emerging from morphology alone, it does not co-evolve any actuation or control. Auerbach and Bongard (2010a) then evolved the placement of CPG controlled rotational joints between cellular spheres, thus co-evolving morphology and control. Cheney et al. (2013) evolved locomoting soft robots made of multiple different materials: two passive voxels of differing rigidity and two actuated voxel types that expanded cyclically via out-of-phase central pattern generators (CPGs). While this work added a variety of soft materials and a new type of actuation, the pairing of muscle types directly to a CPG again reflected a focus on evolving morphology rather than sophisticated neural control. Many examples in the literature include the co-evolution of a robot morphology with an artificial neural network controller [Sims (1994); Lipson and Pollack (2000); Hornby et al. (2001); Lehman and Stanley (2011)]. These studies (and many more like them) involve what might be called “ghost” networks: artificial neural networks that provide control to the body, yet do not have any physical embodiALIFE 14: Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living Systems Figure 2: An example of complex electrical wave propagation in cardiac modeling [Fenton et al. (2005)]. ment in the system they control. The state of input nodes to these networks is often set by sensors in the robot and output nodes typically signify behavioral outcomes in the actuators, but the computation is done supernaturally, disjoint from the body itself. In the age of 3D printing, it is a realistic goal for robots to physically walk out of a printer. It is thus worthwhile to consider designing robots that can be physically realized: i.e., those whose controllers are accounted for by being physically woven into the design of the robot. While the brains of animals are often a separate module within their bodies, animals also have central and peripheral nervous systems extending throughout their bodies. An extreme example of this is the octopus, which has as much as 90% of its neurons existing outside of its central nervous system [Zullo et al. (2009)]. The distributed and physical layout of the nervous system over space may contribute significantly to neural processing, as the delays and branching in axons (the basis for nerves) are suggested to serve computational functions [Segev and Schneidman (1999)]. Despite the prevalence of embodied, distributed circuitry in nearly all of animal life, the idea of an embodied nervous system has been absent from the field of evolutionary robotics. The sub-field called Evolvable Hardware evolves physical circuits for computer chips [Floreano and Mattiussi (2008)], but such work has not been applied to evolving the circuitry of artificial life organisms. We are unaware of work with virtual creatures that have physically embodied control systems (e.g. where neural circuitry physically runs throughout the body of the creature). We present the first such work in this paper. We propose a very basic model of electrical signal propagation throughout the body of an evolved creature. This embodied controller is based on electrophysiology (specifically at large scales, such as cardiac electrophysiology, Fig. 2). Electrophysiology is the study of the electrical properties of biological cells and tissues [Hoffman et al. (1960)]. In this model, electrical pulses from a single centralized sinusoidal pacemaker (analogous to the sinoatrial node – the pacemaker in the heart [Brown (1982)]) are propagated through the electrically conductive tissue of the creature. The location and patterning of this conductive tissue is described by an evolved Compositional Pattern Producing Network (CPPN) genome. Evolution controls the shape of the body and the electrical pathways within it, which both combine to determine the robot’s behavior. The model involves conductive tissue cells that collect voltage from neighboring cells, causing an action potential (spike) if the collected voltage exceeds the cell’s firing threshold (Fig. 3). Once this threshold is crossed, the cell depolarizes, causing a voltage spike that excites neighboring cells. This voltage spike is followed by a refractory period, during which the cell is temporarily unable to be re-excited. This model allows for the propagation of information through the body of the creature in the form of electrical signals. The structure of this flow is produced entirely by the topology of the creature and the state of each cell’s direct neighbors. In this sense, the model can be seen as a form of distributed information processing. One could draw similarities between this model and a 3D-grid of neurons, where each neuron receives inputs from, and has outputs to, its immediate neighbors. In this analogy, we are evolving where neurons should exist in the grid, what type of material the neuron is housed in, as well as the material type, if any, of grid locations that do not contain neurons. The placement of material, which is under evolutionary control, directly determines the resultant behavior of the organism. Cells that actuate will contract and expand as they depolarize (much like the contraction of cardiac muscles), leading to the locomotion behavior of the creature. In order to control the signal flow throughout the creature, insulator cells are allowed, which are unable to accept and pass on the signal. Evolution can also choose not to fill a voxel with material. The morphology of the simulated robot and tissue type at each cell is determined by a CPPN genome. This model examines the evolution of embodied cognition at a more detailed level of implementation than is typical in the literature – with embodied control circuitry resulting directly from the morphology of the individual creature. While this study only covers the classic problem of locomotion, it is a step towards truly physically embodied robots.

A long-standing, open question in biology is how populations are capable of rapidly adapting to novel environments, a trait called evolvability. A major contributor to evolvability is the fact that many biological entities are modular, especially the many biological processes and structures that can be modeled as networks, such as metabolic pathways, gene regulation, protein interactions, and animal brains. Networks are modular if they contain highly connected clusters of nodes that are sparsely connected to nodes in other clusters [4, 2]. Despite its importance and decades of research, there is no agreement on why modularity evolves [4]. Intuitively, modular systems seem more adaptable, a lesson well-known to human engineers, because it is easier to rewire a modular network with functional subunits than an entangled, monolithic network [1]. However, because this evolvability only provides a selective advantage over the long-term, such selection is at best indirect and may not be strong enough to explain the level of modularity in the natural world [4]. Modularity is likely caused by multiple forces acting to various degrees in different contexts [4], and a comprehensive understanding of the evolutionary origins of modularity involves identifying those multiple forces and their relative contributions. The leading hypothesis is that modularity mainly emerges due to rapidly changing environments that have common subproblems, but different overall problems [1]. It is unknown how much natural modularity MVG can explain, however, because it unclear if biological environments change modularly, and whether they change at a high enough frequency for this force to play a significant role. We investigate an alternate hypothesis that has been suggested, but heretofore untested, which is that modularity evolves not because it conveys evolvability, but as a byproduct from selection to reduce connection costs in a network [3].

Abstract Algorithms for evolving agents that learn during their lifetime have typically been evaluated on only a handful of environments. Designing such environments is labour intensive, potentially biased, and provides only a small sample size that may prevent accurate general conclusions from being drawn. In this paper we introduce a method for automatically generating MDP environments which allows the difficulty to be scaled in several ways. We present a case study in which environments are generated that vary along three key dimensions of difficulty: the number of environment configurations, the number of available actions, and the length of each trial. The study reveals interesting differences between three neural network models -- Fixed-Weight, Plastic-Weight, and Modulated Plasticity -- that would not have been obvious without sweeping across these different dimensions. Our paper thus introduces a new way of conducting reinforcement learning science: instead of manually designing a few environments, researchers will be able to automatically generate a range of environments across key dimensions of variation. This will allow scientists to more rigorously assess the general learning capabilities of an algorithm, and may ultimately improve the rate at which we discover how to create AI with general purpose learning.

One of humanity's grand scientific challenges is to create artificially intelligent robots that rival natural animals in intelligence and agility. A key enabler of such animal complexity is the fact that animal brains are structurally organized in that they exhibit modularity and regularity, amongst other attributes. Modularity is the localization of function within an encapsulated unit. Regularity refers to the compressibility of the information describing a structure, and typically involves symmetries and repetition. These properties improve evolvability, but they rarely emerge in evolutionary algorithms without specific techniques to encourage them. It has been shown that (1) modularity can be evolved in neural networks by adding a cost for neural connections and, separately, (2) that the HyperNEAT algorithm produces neural networks with complex, functional regularities. In this paper we show that adding the connection cost technique to HyperNEAT produces neural networks that are significantly more modular, regular, and higher performing than HyperNEAT without a connection cost, even when compared to a variant of HyperNEAT that was specifically designed to encourage modularity. Our results represent a stepping stone towards the goal of producing artificial neural networks that share key organizational properties with the brains of natural animals.

Evolutionary algorithms frequently get stuck on local optima--and fail to find the global optimum--when local gradients do not point the search process toward the direction of the global optimum. A recent breakthrough called Novelty Search ameliorates this problem by enabling the search process to explore in every direction by encouraging the production of novel, or not-yet-seen, phenotypes (e.g. new robot behaviors). However, a problem with Novelty Search is that it can get lost on "novelty plateaus" wherein novel behaviors in offspring are not immediately produced by mutation and crossover (e.g. when a sequence of specific mutations is required to produce new behaviors, but the intermediate mutations are not rewarded because they do not produce novel behaviors). In such cases, Novelty Search and related approaches that reward behavioral diversity can get stuck. Here we introduce a new approach, borrowed from human psychology, that mitigates this problem: encouraging creative thinking. In addition to rewarding novel behavior, we encourage evolving neural networks to "think differently" by rewarding not-yet-seen firing patterns in hidden neurons, which we call the "Creative Thinking Approach." We hypothesize that encouraging novel thinking can reward stepping stones toward new behaviors. On a variety of challenging robotic control problems from previous publications we demonstrate that, as problem difficulty increases, adding the Creative Thinking Approach increasingly improves performance over simply encouraging novel behaviors. Our results suggest that the Creative Thinking Approach could help improve the scale and complexity of problems that can be solved by evolutionary algorithms.

Novelty Search, a new type of Evolutionary Algorithm, has shown much promise in the last few years. Instead of selecting for phenotypes that are closer to an objective, Novelty Search assigns rewards based on how different the phenotypes are from those already generated. A common criticism of Novelty Search is that it is effectively random or exhaustive search because it tries solutions in an unordered manner until a correct one is found. Its creators respond that over time Novelty Search accumulates information about the environment in the form of skills relevant to reaching uncharted territory, but to date no evidence for that hypothesis has been presented. In this paper we test that hypothesis by transferring robots evolved under Novelty Search to new environments (here, mazes) to see if the skills they've acquired generalize. Three lines of evidence support the claim that Novelty Search agents do indeed learn general exploration skills. First, robot controllers evolved via Novelty Search in one maze and then transferred to a new maze explore significantly more of the new environment than non-evolved (randomly generated) agents. Second, a Novelty Search process to solve the new mazes works significantly faster when seeded with the transferred controllers versus randomly-generated ones. Third, no significant difference exists when comparing two types of transferred agents: those evolved in the original maze under (1) Novelty Search vs. (2) a traditional, objective-based fitness function. The evidence gathered suggests that, like traditional Evolutionary Algorithms with objective-based fitness functions, Novelty Search is not a random or exhaustive search process, but instead is accumulating information about the environment, resulting in phenotypes possessing skills needed to explore their world.

Ongoing, rapid advances in three-dimensional (3D) printing technology are making it inexpensive for lay people to manufacture 3D objects. However, the lack of tools to help nontechnical users design interesting, complex objects represents a significant barrier preventing the public from benefitting from 3D printers. Previous work has shown that an evolutionary algorithm with a generative encoding based on developmental biology-a compositional pattern-producing network (CPPN)-can automate the design of interesting 3D shapes, but users collectively had to start each act of creation from a random object, making it difficult to evolve preconceived target shapes. In this paper, we describe how to modify that algorithm to allow the further evolution of any uploaded shape. The technical insight is to inject the distance to the surface of the object as an input to the CPPN. We show that this seeded-CPPN technique reproduces the original shape to an arbitrary resolution, yet enables morphing the shape in interesting, complex ways. This technology also raises the possibility of two new, important types of science: (1) It could work equally well for CPPN-encoded neural networks, meaning neural wiring diagrams from nature, such as the mouse or human connectome, could be injected into a neural network and further evolved via the CPPN encoding. (2) The technique could be generalized to recreate any CPPN phenotype, but substituting a flat CPPN representation for the rich, originally evolved one. Any evolvability extant in the original CPPN genome can be assessed by comparing the two, a project we take first steps toward in this paper. Overall, this paper introduces a method that will enable non-technical users to modify complex, existing 3D shapes and opens new types of scientific inquiry that can catalyze research on bio-inspired artificial intelligence and the evolvability benefits of generative encodings.

In 1994 Karl Sims showed that computational evolution can produce interesting morphologies that resemble natural organisms. Despite nearly two decades of work since, evolved morphologies are not obviously more complex or natural, and the field seems to have hit a complexity ceiling. One hypothesis for the lack of increased complexity is that most work, including Sims', evolves morphologies composed of rigid elements, such as solid cubes and cylinders, limiting the design space. A second hypothesis is that the encodings of previous work have been overly regular, not allowing complex regularities with variation. Here we test both hypotheses by evolving soft robots with multiple materials and a powerful generative encoding called a compositional pattern-producing network (CPPN). Robots are selected for locomotion speed. We find that CPPNs evolve faster robots than a direct encoding and that the CPPN morphologies appear more natural. We also find that locomotion performance increases as more materials are added, that diversity of form and behavior can be increased with different cost functions without stifling performance, and that organisms can be evolved at different levels of resolution. These findings suggest the ability of generative soft-voxel systems to scale towards evolving a large diversity of complex, natural, multi-material creatures. Our results suggest that future work that combines the evolution of CPPN-encoded soft, multi-material robots with modern diversity-encouraging techniques could finally enable the creation of creatures far more complex and interesting than those produced by Sims nearly twenty years ago.

We describe a new, quadruped robot platform, Aracna, which requires non-intuitive motor commands in order to locomote and thus provides an interesting challenge for gait learning algorithms, such as those frequently developed in the Evolu-tionary Computation and Artificial Life communities. Aracna is an open-source hardware project composed of off-the-shelf and 3D-printed parts, enabling other research teams to mod-ify its design according to their scientific needs. Aracna was designed to overcome the shortcomings of a previous quadruped robot platform, whose legs were so heavy that the motors could not reliably execute the commands sent to them. We avoid this problem by locating all motors in the body core instead of on the legs and through a design which enables the servos to have a greater mechanical advantage. Specifically, each of the four legs has two joints controlled by separate four-bar linkage mechanisms that drive the pitch of the hip joint and knee joint. This novel design causes unconventional kinematics, creating an opportunity for gait-learning algorithms, which excel in counter-intuitive design spaces where human engineers tend to underperform. Be-cause it is low-cost, flexible, kinematically interesting, and and improvement over a previous design, Aracna provides a useful new hardware platform for testing algorithms that au-tomatically generate robotic behaviors.

In this paper we introduce the Developmental Symbolic Encoding (DSE), a new generative encoding for evolving networks (e.g. neural or boolean). DSE combines elements of two powerful generative encodings, Cellular Encoding and HyperNEAT, in order to evolve networks that are modular, regular, scale-free, and scalable. Generating networks with these properties is important because they can enhance performance and evolvability. We test DSE's ability to generate scale-free and modular networks by explicitly rewarding these properties and seeing whether evolution can produce networks that possess them. We compare the networks DSE evolves to those of HyperNEAT. The results show that both encodings can produce scale-free networks, although DSE performs slightly, but significantly, better on this objective. DSE networks are far more modular than HyperNEAT networks. Both encodings produce regular networks. We further demonstrate that individual DSE genomes during development can scale up a network pattern to accommodate different numbers of inputs. We also compare DSE to HyperNEAT on a pattern recognition problem. DSE significantly outperforms HyperNEAT, suggesting that its potential lay not just in the properties of the networks it produces, but also because it can compete with leading encodings at solving challenging problems. These preliminary results imply that DSE is an interesting new encoding worthy of additional study. The results also raise questions about which network properties are more likely to be produced by different types of generative encodings.

This paper investigates how an evolutionary algorithm with an indirect encoding exploits the property of phenotypic regularity, an important design principle found in natural organisms and engineered designs. We present the first comprehensive study showing that such phenotypic regularity enables an indirect encoding to outperform direct encoding controls as problem regularity increases. Such an ability to produce regular solutions that can exploit the regularity of problems is an important prerequisite if evolutionary algorithms are to scale to high-dimensional real-world problems, which typically contain many regularities, both known and unrecognized. The indirect encoding in this case study is HyperNEAT, which evolves artificial neural networks (ANNs) in a manner inspired by concepts from biological development. We demonstrate that, in contrast to two direct encoding controls, HyperNEAT produces both regular behaviors and regular ANNs, which enables HyperNEAT to significantly outperform the direct encodings as regularity increases in three problem domains. We also show that the types of regularities HyperNEAT produces can be biased, allowing domain knowledge and preferences to be injected into the search. Finally, we examine the downside of a bias toward regularity. Even when a solution is mainly regular, some irregularity may be needed to perfect its functionality. This insight is illustrated by a new algorithm called HybrID that hybridizes indirect and direct encodings, which matched HyperNEAT's performance on regular problems yet outperformed it on problems with some irregularity. HybrID's ability to improve upon the performance of HyperNEAT raises the question of whether indirect encodings may ultimately excel not as stand-alone algorithms, but by being hybridized with a further process of refinement, wherein the indirect encoding produces patterns that exploit problem regularity and the refining process modifies that pattern to capture irregularities. This paper thus paints a more complete picture of indirect encodings than prior studies because it analyzes the impact of the continuum between irregularity and regularity on the performance of such encodings, and ultimately suggests a path forward that combines indirect encodings with a separate process of refinement.

HyperNEAT represents a class of neuroevolutionary algorithms that captures some of the power of natural development with a computationally efficient high-level abstraction of development. This class of algorithms is intended to provide many of the desirable properties produced in biological phenotypes by natural developmental processes, such as regularity, modularity and hierarchy. While it has been previously shown that HyperNEAT produces regular artificial neural network (ANN) phenotypes, in this paper we investigated the open question of whether HyperNEAT can produce modular ANNs. We conducted such research on problems where modularity should be beneficial, and found that HyperNEAT failed to generate modular ANNs. We then imposed modularity on HyperNEAT’s phenotypes and its performance improved, demonstrating that modularity increases performance on this problem. We next tested two techniques to encourage modularity in HyperNEAT, but did not observe an increase in either modularity or performance. Finally, we conducted tests on a simpler problem that requires modularity and found that HyperNEAT was able to rapidly produce modular solutions that solved the problem. We therefore present the first documented case of HyperNEAT producing a modular phenotype, but our inability to encourage modularity on harder problems where modularity would have been beneficial suggests that more work is needed to increase the likelihood that HyperNEAT and similar algorithms produce modular ANNs in response to challenging, decomposable problems.

Legged robots show promise for complex mobility tasks, such as navigating rough terrain, but the design of their control software is both challenging and laborious. Traditional evolutionary algorithms can produce these controllers, but require manual decomposition or other problem simplification because conventionally-used direct encodings have trouble taking advantage of a problem's regularities and symmetries. Such active intervention is time consuming, limits the range of potential solutions, and requires the user to possess a deep understanding of the problem's structure. This paper demonstrates that HyperNEAT, a new and promising generative encoding for evolving neural networks, can evolve quadruped gaits without an engineer manually decomposing the problem. Analyses suggest that HyperNEAT is successful because it employs a generative encoding that can more easily reuse phenotypic modules. It is also one of the first neuroevolutionary algorithms that exploits a problem's geometric symmetries, which may aid its performance. We compare HyperNEAT to FT-NEAT, a direct encoding control, and find that HyperNEAT is able to evolve impressive quadruped gaits and vastly outperforms FT-NEAT. Comparative analyses reveal that HyperNEAT individuals are more holistically affected by genetic operators, resulting in better leg coordination. Overall, the results suggest that HyperNEAT is a powerful algorithm for evolving control systems for complex, yet regular, devices, such as robots.

HyperNEAT, a generative encoding for evolving artificial neural networks (ANNs), has the unique and powerful ability to exploit the geometry of a problem (e.g., symmetries) by encoding ANNs as a function of a problem's geometry. This paper provides the first extensive analysis of the sensitivity of HyperNEAT to different geometric representations of a problem. Understanding how geometric representations affect the quality of evolved solutions should improve future designs of such representations. HyperNEAT has been shown to produce coordinated gaits for a simulated quadruped robot with a specific two-dimensional geometric representation. Here, the same problem domain is tested, but with different geometric representations of the problem. Overall, experiments show that the quality and kind of solutions produced by HyperNEAT can be substantially affected by the geometric representation. HyperNEAT outperforms a direct encoding control even with randomized geometric representations, but performs even better when a human engineer designs a representation that reflects the actual geometry of the robot. Unfortunately, even choices in geometric layout that seem to be inconsequential a priori can significantly affect fitness. Additionally, a geometric representation can bias the type of solutions generated (e.g., make left-right symmetry more common than front-back symmetry). The results suggest that HyperNEAT practitioners can obtain good results even if they do not know how to geometrically represent a problem, and that further improvements are possible with a well-chosen geometric representation.

All in-text\treferences\tunderlined\tin\tblue\tare\tlinked\tto\tpublications\ton\tResearchGate, letting you\taccess\tand\tread\tthem\timmediately.

Generative representations allow the reuse of code and thus facilitate the evolution of repeated phenotypic themes or modules. It has been shown that generative representations perform well on highly regular problems. To date, however, generative representations have not been tested on irregular problems. It is unknown how fast their performance degrades as the regularity of the problem decreases. In this report, we test a generative representation on a problem where we can scale a type of regularity in the problem. The generative representation outperforms a direct encoding control when the regularity of the problem is high but degrades to, and then underperforms, the direct control as the regularity of the problem decreases. Importantly, this decrease is not linear. The boost provided by the generative encoding is only significant for very high levels of regularity.

The rate of mutation is central to evolution.Mutations are required for adaptation, yet most mutations with phenotypic effects are deleterious.As a consequence, the mutation rate that maximizes adaptation will be some intermediate value.Here, we used digital organisms to investigate the ability of natural selection to adjust and optimize mutation rates.We assessed the optimal mutation rate by empirically determining what mutation rate produced the highest rate of adaptation.Then, we allowed mutation rates to evolve, and we evaluated the proximity to the optimum.Although we chose conditions favorable for mutation rate optimization, the evolved rates were invariably far below the optimum across a wide range of experimental parameter settings.We hypothesized that the reason that mutation rates evolved to be suboptimal was the ruggedness of fitness landscapes.To test this hypothesis, we created a simplified landscape without any fitness valleys and found that, in such conditions, populations evolved near-optimal mutation rates.In contrast, when fitness valleys were added to this simple landscape, the ability of evolving populations to find the optimal mutation rate was lost.We conclude that rugged fitness landscapes can prevent the evolution of mutation rates that are optimal for long-term adaptation.This finding has important implications for applied evolutionary research in both biological and computational realms.