Scaling up inference-time compute has proven to be a valuable strategy in improving the performance of Large Language Models (LLMs) without fine-tuning. An important task that can benefit from additional inference-time compute is self-repair; given an initial flawed response or guess, the LLM corrects its own mistake and produces an improved response or fix. We leverage the in-context learning ability of LLMs to perform self-repair in the coding domain. The key contribution of our paper is an approach that synthesises and selects an ordered set of golden example pairs, or AuPairs, of these initial guesses and subsequent fixes for the corresponding problems. Each such AuPair is provided as a single in-context example at inference time to generate a repaired solution. For an inference-time compute budget of $N$ LLM calls per problem, $N$ AuPairs are used to generate $N$ repaired solutions, out of which the highest-scoring solution is the final answer. The underlying intuition is that if the LLM is given a different example of fixing an incorrect guess each time, it can subsequently generate a diverse set of repaired solutions. Our algorithm selects these AuPairs in a manner that maximises complementarity and usefulness. We demonstrate the results of our algorithm on 5 LLMs across 7 competitive programming datasets for the code repair task. Our algorithm yields a significant boost in performance compared to best-of-$N$ and self-repair, and also exhibits strong generalisation across datasets and models. Moreover, our approach shows stronger scaling with inference-time compute budget compared to baselines.

Agents are minimally entities that are influenced by their past observations and act to influence future observations. This latter capacity is captured by empowerment, which has served as a vital framing concept across artificial intelligence and cognitive science. This former capacity, however, is equally foundational: In what ways, and to what extent, can an agent be influenced by what it observes? In this paper, we ground this concept in a universal agent-centric measure that we refer to as plasticity, and reveal a fundamental connection to empowerment. Following a set of desiderata on a suitable definition, we define plasticity using a new information-theoretic quantity we call the generalized directed information. We show that this new quantity strictly generalizes the directed information introduced by Massey (1990) while preserving all of its desirable properties. Under this definition, we find that plasticity is well thought of as the mirror of empowerment: The two concepts are defined using the same measure, with only the direction of influence reversed. Our main result establishes a tension between the plasticity and empowerment of an agent, suggesting that agent design needs to be mindful of both characteristics. We explore the implications of these findings, and suggest that plasticity, empowerment, and their relationship are essential to understanding agency.

The quality of foundation models depends heavily on their training data. Consequently, great efforts have been put into dataset curation. Yet most approaches rely on manual tuning of coarse-grained mixtures of large buckets of data, or filtering by hand-crafted heuristics. An approach that is ultimately more scalable (let alone more satisfying) is to \emph{learn} which data is actually valuable for training. This type of meta-learning could allow more sophisticated, fine-grained, and effective curation. Our proposed \emph{DataRater} is an instance of this idea. It estimates the value of training on any particular data point. This is done by meta-learning using `meta-gradients', with the objective of improving training efficiency on held out data. In extensive experiments across a range of model scales and datasets, we find that using our DataRater to filter data is highly effective, resulting in significantly improved compute efficiency.

In recent years there has been a tremendous surge in the general capabilities of AI systems, mainly fuelled by training foundation models on internet-scale data. Nevertheless, the creation of open-ended, ever self-improving AI remains elusive. **In this position paper, we argue that the ingredients are now in place to achieve *open-endedness* in AI systems with respect to a human observer. Furthermore, we claim that such open-endedness is an essential property of any artificial superhuman intelligence (ASI).** We begin by providing a concrete formal definition of open-endedness through the lens of novelty and learnability. We then illustrate a path towards ASI via open-ended systems built on top of foundation models, capable of making novel, human-relevant discoveries. We conclude by examining the safety implications of generally-capable open-ended AI. We expect that open-ended foundation models will prove to be an increasingly fertile and safety-critical area of research in the near future.

Genetic algorithms constitute a family of black-box optimization algorithms, which take inspiration from the principles of biological evolution. While they provide a general-purpose tool for optimization, their particular instantiations can be heuristic and motivated by loose biological intuition. In this work we explore a fundamentally different approach: Given a sufficiently flexible parametrization of the genetic operators, we discover entirely new genetic algorithms in a data-driven fashion. More specifically, we parametrize selection and mutation rate adaptation as cross- and self-attention modules and use Meta-Black-Box-Optimization to evolve their parameters on a set of diverse optimization tasks. The resulting Learned Genetic Algorithm outperforms state-of-the-art adaptive baseline genetic algorithms and generalizes far beyond its meta-training settings. The learned algorithm can be applied to previously unseen optimization problems, search dimensions & evaluation budgets. We conduct extensive analysis of the discovered operators and provide ablation experiments, which highlight the benefits of flexible module parametrization and the ability to transfer ('plug-in') the learned operators to conventional genetic algorithms.

One of the gnarliest challenges in reinforcement learning (RL) is exploration that scales to vast domains, where novelty-, or coverage-seeking behaviour falls short. Goal-directed, purposeful behaviours are able to overcome this, but rely on a good goal space. The core challenge in goal discovery is finding the right balance between generality (not hand-crafted) and tractability (useful, not too many). Our approach explicitly seeks the middle ground, enabling the human designer to specify a vast but meaningful proto-goal space, and an autonomous discovery process to refine this to a narrower space of controllable, reachable, novel, and relevant goals. The effectiveness of goal-conditioned exploration with the latter is then demonstrated in three challenging environments.

Using a model of the environment and a value function, an agent can construct many estimates of a state’s value, by unrolling the model for different lengths and bootstrapping with its value function. Our key insight is that one can treat this set of value estimates as a type of ensemble, which we call an implicit value ensemble (IVE). Consequently, the discrepancy between these estimates can be used as a proxy for the agent’s epistemic uncertainty; we term this signal model-value inconsistency or self-inconsistency for short. Unlike prior work which estimates uncertainty by training an ensemble of many models and/or value functions, this approach requires only the single model and value function which are already being learned in most model-based reinforcement learning algorithms. We provide empirical evidence in both tabular and function approximation settings from pixels that self-inconsistency is useful (i) as a signal for exploration, (ii) for acting safely under distribution shifts, and (iii) for robustifying value-based planning with a learned model.

We identify and study the phenomenon of policy churn, that is, the rapid change of the greedy policy in value-based reinforcement learning. Policy churn operates at a surprisingly rapid pace, changing the greedy action in a large fraction of states within a handful of learning updates (in a typical deep RL set-up such as DQN on Atari). We characterise the phenomenon empirically, verifying that it is not limited to specific algorithm or environment properties. A number of ablations help whittle down the plausible explanations on why churn occurs to just a handful, all related to deep learning. Finally, we hypothesise that policy churn is a beneficial but overlooked form of implicit exploration that casts $\epsilon$-greedy exploration in a fresh light, namely that $\epsilon$-noise plays a much smaller role than expected.

The principal contribution of this paper is a conceptual framework for off-policy reinforcement learning, based on conditional expectations of importance sampling ratios. This framework yields new perspectives and understanding of existing off-policy algorithms, and reveals a broad space of unexplored algorithms. We theoretically analyse this space, and concretely investigate several algorithms that arise from this framework.

The deep reinforcement learning community has made several independent improvements to the DQN algorithm. However, it is unclear which of these extensions are complementary and can be fruitfully combined. This paper examines six extensions to the DQN algorithm and empirically studies their combination. Our experiments show that the combination provides state-of-the-art performance on the Atari 2600 benchmark, both in terms of data efficiency and final performance. We also provide results from a detailed ablation study that shows the contribution of each component to overall performance.

Deep reinforcement learning (RL) has achieved several high profile successes in difficult decision-making problems. However, these algorithms typically require a huge amount of data before they reach reasonable performance. In fact, their performance during learning can be extremely poor. This may be acceptable for a simulator, but it severely limits the applicability of deep RL to many real-world tasks, where the agent must learn in the real environment. In this paper we study a setting where the agent may access data from previous control of the system. We present an algorithm, Deep Q-learning from Demonstrations (DQfD), that leverages small sets of demonstration data to massively accelerate the learning process even from relatively small amounts of demonstration data and is able to automatically assess the necessary ratio of demonstration data while learning thanks to a prioritized replay mechanism. DQfD works by combining temporal difference updates with supervised classification of the demonstrator’s actions. We show that DQfD has better initial performance than Prioritized Dueling Double Deep Q-Networks (PDD DQN) as it starts with better scores on the first million steps on 41 of 42 games and on average it takes PDD DQN 83 million steps to catch up to DQfD’s performance. DQfD learns to out-perform the best demonstration given in 14 of 42 games. In addition, DQfD leverages human demonstrations to achieve state-of-the-art results for 11 games. Finally, we show that DQfD performs better than three related algorithms for incorporating demonstration data into DQN.

The ability to transfer skills across tasks has the potential to scale up reinforcement learning (RL) agents to environments currently out of reach. Recently, a framework based on two ideas, successor features (SFs) and generalised policy improvement (GPI), has been introduced as a principled way of transferring skills. In this paper we extend the SF&GPI framework in two ways. One of the basic assumptions underlying the original formulation of SF&GPI is that rewards for all tasks of interest can be computed as linear combinations of a fixed set of features. We relax this constraint and show that the theoretical guarantees supporting the framework can be extended to any set of tasks that only differ in the reward function. Our second contribution is to show that one can use the reward functions themselves as features for future tasks, without any loss of expressiveness, thus removing the need to specify a set of features beforehand. This makes it possible to combine SF&GPI with deep learning in a more stable way. We empirically verify this claim on a complex 3D environment where observations are images from a first-person perspective. We show that the transfer promoted by SF&GPI leads to very good policies on unseen tasks almost instantaneously. We also describe how to learn policies specialised to the new tasks in a way that allows them to be added to the agent’s set of skills, and thus be reused in the future.

One of the key challenges of artificial intelligence is to learn models that are effective in the context of planning. In this document we introduce the predictron architecture. The predictron consists of a fully abstract model, represented by a Markov reward process, that can be rolled forward multiple “imagined” planning steps. Each forward pass of the predictron accumulates internal rewards and values over multiple planning depths. The predictron is trained end-to-end so as to make these accumulated values accurately approximate the true value function. We applied the predictron to procedurally generated random mazes and a simulator for the game of pool. The predictron yielded significantly more accurate predictions than conventional deep neural network architectures.

We introduce FeUdal Networks (FuNs): a novel architecture for hierarchical reinforcement learning. Our approach is inspired by the feudal reinforcement learning proposal of Dayan and Hinton, and gains power and efficacy by decoupling end-to-end learning across multiple levels – allowing it to utilise different resolutions of time. Our framework employs a Manager module and a Worker module. The Manager operates at a slower time scale and sets abstract goals which are conveyed to and enacted by the Worker. The Worker generates primitive actions at every tick of the environment. The decoupled structure of FuN conveys several benefits – in addition to facilitating very long timescale credit assignment it also encourages the emergence of sub-policies associated with different goals set by the Manager. These properties allow FuN to dramatically outperform a strong baseline agent on tasks that involve long-term credit assignment or memorisation.

Transfer in reinforcement learning refers to the notion that generalization should occur not only within a task but also across tasks. We propose a transfer framework for the scenario where the reward function changes between tasks but the environment's dynamics remain the same. Our approach rests on two key ideas: "successor features", a value function representation that decouples the dynamics of the environment from the rewards, and "generalized policy improvement", a generalization of dynamic programming's policy improvement operation that considers a set of policies rather than a single one. Put together, the two ideas lead to an approach that integrates seamlessly within the reinforcement learning framework and allows the free exchange of information across tasks. The proposed method also provides performance guarantees for the transferred policy even before any learning has taken place. We derive two theorems that set our approach in firm theoretical ground and present experiments that show that it successfully promotes transfer in practice, significantly outperforming alternative methods in a sequence of navigation tasks and in the control of a simulated robotic arm.

Neural networks have a smooth initial inductive bias, such that small changes in input do not lead to large changes in output. However, in reinforcement learning domains with sparse rewards, value functions have non-smooth structure with a characteristic asymmetric discontinuity whenever rewards arrive. We propose a mechanism that learns an interpolation between a direct value estimate and a projected value estimate computed from the encountered reward and the previous estimate. This reduces the need to learn about discontinuities, and thus improves the value function approximation. Furthermore, as the interpolation is learned and state-dependent, our method can deal with heterogeneous observability. We demonstrate that this one change leads to significant improvements on multiple Atari games, when applied to the state-of-the-art A3C algorithm.

The General Video Game AI framework and competition pose the problem of creating artificial intelligence that can play a wide, and in principle unlimited, range of games. Concretely, it tackles the problem of devising an algorithm that is able to play any game it is given, even if the game is not known a priori. This area of study can be seen as an approximation of General Artificial Intelligence, with very little room for game-dependent heuristics. This short paper summarizes the motivation, infrastructure, results and future plans of General Video Game AI, stressing the findings and first conclusions drawn after two editions of our competition, and outlining our future plans.

In recent years there have been many successes of using deep representations in reinforcement learning. Still, many of these applications use conventional architectures, such as convolutional networks, LSTMs, or auto-encoders. In this paper, we present a new neural network architecture for model-free reinforcement learning. Our dueling network represents two separate estimators: one for the state value function and one for the state-dependent action advantage function. The main benefit of this factoring is to generalize learning across actions without imposing any change to the underlying reinforcement learning algorithm. Our results show that this architecture leads to better policy evaluation in the presence of many similar-valued actions. Moreover, the dueling architecture enables our RL agent to outperform the state-of-the-art on the Atari 2600 domain.

The move from hand-designed features to learned features in machine learning has been wildly successful. In spite of this, optimization algorithms are still designed by hand. In this paper we show how the design of an optimization algorithm can be cast as a learning problem, allowing the algorithm to learn to exploit structure in the problems of interest in an automatic way. Our learned algorithms, implemented by LSTMs, outperform generic, hand-designed competitors on the tasks for which they are trained, and also generalize well to new tasks with similar structure. We demonstrate this on a number of tasks, including simple convex problems, training neural networks, and styling images with neural art.

We consider an agent's uncertainty about its environment and the problem of generalizing this uncertainty across states. Specifically, we focus on the problem of exploration in non-tabular reinforcement learning. Drawing inspiration from the intrinsic motivation literature, we use density models to measure uncertainty, and propose a novel algorithm for deriving a pseudo-count from an arbitrary density model. This technique enables us to generalize count-based exploration algorithms to the non-tabular case. We apply our ideas to Atari 2600 games, providing sensible pseudo-counts from raw pixels. We transform these pseudo-counts into exploration bonuses and obtain significantly improved exploration in a number of hard games, including the infamously difficult Montezuma's Revenge.

This paper presents the framework, rules, games, controllers, and results of the first General Video Game Playing Competition, held at the IEEE Conference on Computational Intelligence and Games in 2014. The competition proposes the challenge of creating controllers for general video game play, where a single agent must be able to play many different games, some of them unknown to the participants at the time of submitting their entries. This test can be seen as an approximation of general artificial intelligence, as the amount of game-dependent heuristics needs to be severely limited. The games employed are stochastic real-time scenarios (where the time budget to provide the next action is measured in milliseconds) with different winning conditions, scoring mechanisms, sprite types, and available actions for the player. It is a responsibility of the agents to discover the mechanics of each game, the requirements to obtain a high score and the requisites to finally achieve victory. This paper describes all controllers submitted to the competition, with an in-depth description of four of them by their authors, including the winner and the runner-up entries of the contest. The paper also analyzes the performance of the different approaches submitted, and finally proposes future tracks for the competition.

Value functions are a core component of reinforcement learning. The main idea is to to construct a single function approximator V(s; theta) that estimates the long-term reward from any state s, using parameters θ. In this paper we introduce universal value function approximators (UVFAs) V(s,g;theta) that generalise not just over states s but also over goals g. We develop an efficient technique for supervised learning of UVFAs, by factoring observed values into separate embedding vectors for state and goal, and then learning a mapping from s and g to these factored embedding vectors. We show how this technique may be incorporated into a reinforcement learning algorithm that updates the UVFA solely from observed rewards. Finally, we demonstrate that a UVFA can successfully generalise to previously unseen goals.

In this short paper, we propose a powerful new tool for conducting research on computational intelligence and games. “PyVGDL” is a simple, high-level, extensible description language for 2-D video games. It is based on defining locations and dynamics for simple building blocks (objects), together with local interaction effects. A rich ontology defines various controllers, object behaviors, passive effects (physics), and collision effects. It can be used to quickly design games, without having to deal with control structures. We show how the dynamics of many classical games can be generated from a few lines of PyVGDL. Furthermore, the accompanying software library permits parsing and instantly playing those games, visualized from a bird's-eye or first-person viewpoint, and using them as benchmarks for learning algorithms.

The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of the best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

This theoretical investigation gives the first proof of convergence for (radial) natural evolution strategies, on d-dimensional sphere functions, and establishes the conditions on hyper-parameters, as a function of d. For the limit case of large population sizes we show asymptotic linear convergence, and in the limit of small learning rates we give a full analytic characterization of the algorithm dynamics, decomposed into transient and asymptotic phases. Finally, we show why omitting the natural gradient is catastrophic.

The principle of artificial curiosity directs active exploration towards the most informative or most interesting data. We show its usefulness for global black box optimization when data point evaluations are expensive. Gaussian process regression is used to model the fitness function based on all available observations so far. For each candidate point this model estimates expected Fitness reduction, and yields a novel closed-form expression of expected information gain. A new type of Pareto-front algorithm continually pushes the boundary of candidates not dominated by any other known data according to both criteria, using multi-objective evolutionary search. This makes the exploration-exploitation trade-off explicit, and permits maximally informed data selection. We illustrate the robustness of our approach in a number of experimental scenarios.

The family of natural evolution strategies (NES) offers a principled approach to real-valued evolutionary optimization. NES follows the natural gradient of the expected fitness on the parameters of its search distribution. While general in its formulation, previous research has focused on multivariate Gaussian search distributions. Here we exhibit problem classes for which other search distributions are more appropriate, and then derive corresponding NES-variants.

This paper introduces a novel multi-modular method for reinforcement learning. A multimodular system is one that partitions the learning task among a set of experts (modules), where each expert is incapable of solving the entire task by itself. There are many advantages to splitting up large tasks in this way, but existing methods face difficulties when choosing which module(s) should contribute to the agent’s actions at any particular moment. We introduce a novel selection mechanism where every module, besides calculating a set of action values, also estimates its own error for the current input. The selection mechanism combines each module’s estimate of long-term reward and self-error to produce a score by which the next module is chosen. As a result, the modules can use their resources effectively and efficiently divide up the task. The system is shown to learn complex tasks even when the individual modules use only linear function approximators. 1

The family of natural evolution strategies (NES) offers a principled approach to real-valued evolutionary optimization by following the natural gradient of the expected fitness. Like the well-known CMA-ES, the most competitive algorithm in the field, NES comes with important invariance properties. In this paper, we introduce a number of elegant and efficient improvements of the basic NES algorithm. First, we propose to parameterize the positive definite covariance matrix using the exponential map, which allows the covariance matrix to be updated in a vector space. This new technique makes the algorithm completely invariant under linear transformations of the underlying search space, which was previously achieved only in the limit of small step sizes. Second, we compute all updates in the natural coordinate system, such that the natural gradient coincides with the vanilla gradient. This way we avoid the computation of the inverse Fisher information matrix, which is the main computational bottleneck of the original NES algorithm. Our new algorithm, exponential NES (xNES), is significantly simpler than its predecessors. We show that the various update rules in CMA-ES are closely related to the natural gradient updates of xNES. However, xNES is more principled than CMA-ES, as all the update rules needed for covariance matrix adaptation are derived from a single principle. We empirically assess the performance of the new algorithm on standard benchmark functions

Efficient Natural Evolution Strategies (eNES) is a novel alternative to conventional evolutionary algorithms, using the natural gradient to adapt the mutation distribution. Unlike previous methods based on natural gradients, eNES uses a fast algorithm to calculate the inverse of the exact Fisher information matrix, thus increasing both robustness and performance of its evolution gradient estimation, even in higher dimensions. Additional novel aspects of eNES include optimal fitness baselines and importance mixing (a procedure for updating the population with very few fitness evaluations). The algorithm yields competitive results on both unimodal and multimodal benchmarks.

To optimize unknown 'fitness' functions, we present Natural Evolution Strategies, a novel algorithm that constitutes a principled alternative to standard stochastic search methods. It maintains a multinormal distribution on the set of solution candidates. The Natural Gradient is used to update the distribution's parameters in the direction of higher expected fitness, by efficiently calculating the inverse of the exact Fisher information matrix whereas previous methods had to use approximations. Other novel aspects of our method include optimal fitness baselines and importance mixing, a procedure adjusting batches with minimal numbers of fitness evaluations. The algorithm yields competitive results on a number of benchmarks.

This paper presents natural evolution strategies (NES), a novel algorithm for performing real-valued dasiablack boxpsila function optimization: optimizing an unknown objective function where algorithm-selected function measurements constitute the only information accessible to the method. Natural evolution strategies search the fitness landscape using a multivariate normal distribution with a self-adapting mutation matrix to generate correlated mutations in promising regions. NES shares this property with covariance matrix adaption (CMA), an evolution strategy (ES) which has been shown to perform well on a variety of high-precision optimization tasks. The natural evolution strategies algorithm, however, is simpler, less ad-hoc and more principled. Self-adaptation of the mutation matrix is derived using a Monte Carlo estimate of the natural gradient towards better expected fitness. By following the natural gradient instead of the dasiavanillapsila gradient, we can ensure efficient update steps while preventing early convergence due to overly greedy updates, resulting in reduced sensitivity to local suboptima. We show NES has competitive performance with CMA on unimodal tasks, while outperforming it on several multimodal tasks that are rich in deceptive local optima.