Prompting is one of the main ways to adapt a pretrained model to target tasks. Besides manually constructing prompts, many prompt optimization methods have been proposed in the literature. Method development is mainly empirically driven, with less emphasis on a conceptual understanding of prompting. In this paper we discuss how optimal prompting can be understood through a Bayesian view, which also implies some fundamental limitations of prompting that can only be overcome by tuning weights. The paper explains in detail how meta-trained neural networks behave as Bayesian predictors over the pretraining distribution, whose hallmark feature is rapid in-context adaptation. Optimal prompting can be studied formally as conditioning these Bayesian predictors, yielding criteria for target tasks where optimal prompting is and is not possible. We support the theory with educational experiments on LSTMs and Transformers, where we compare different versions of prefix-tuning and different weight-tuning methods. We also confirm that soft prefixes, which are sequences of real-valued vectors outside the token alphabet, can lead to very effective prompts for trained and even untrained networks by manipulating activations in ways that are not achievable by hard tokens. This adds an important mechanistic aspect beyond the conceptual Bayesian theory.

In reinforcement learning (RL), if the agent’s reward differs from the designers’ true utility, even only rarely, the state distribution resulting from the agent’s policy can be very bad, in theory and in practice. When RL policies would devolve into undesired behavior, a common countermeasure is KL regularization to a trusted policy ("Don’t do anything I wouldn’t do"). All current cutting-edge language models are RL agents that are KL-regularized to a "base policy" that is purely predictive. Unfortunately, we demonstrate that when this base policy is a Bayesian predictive model of a trusted policy, the KL constraint is no longer reliable for controlling the behavior of an advanced RL agent. We demonstrate this theoretically using algorithmic information theory, and while systems today are too weak to exhibit this theorized failure precisely, we RL-finetune a language model and find evidence that our formal results are plausibly relevant in practice. We also propose a theoretical alternative that avoids this problem by replacing the "Don’t do anything I wouldn’t do" principle with "Don’t do anything I mightn’t do".

The domain-independent universal Normalized Information Distance based on Kolmogorov complexity has been (in approximate form) successfully applied to a variety of difficult clustering problems. In this paper we investigate theoretical properties of the un-normalized algorithmic information distance$d_{K}$. The main question we are asking in this work is what properties this curious distance has, besides being a metric. We show that many (in)finite-dimensional spaces can(not) be isometrically scale-embedded into the space of finite strings with metric$d_{K}$. We also show that$d_{K}$is not an Euclidean distance, but any finite set of points in Euclidean space can be scale-embedded into$(\{0,1\}^{*},d_{K})$. A major contribution is the development of the necessary framework and tools for finding more (interesting) properties of$d_{K}$in future, and to state several open problems.

Prior approximations of AIXI, a Bayesian optimality notion for general reinforcement learning, can only approximate AIXI's Bayesian environment model using an a-priori defined set of models. This is a fundamental source of epistemic uncertainty for the agent in settings where the existence of systematic bias in the predefined model class cannot be resolved by simply collecting more data from the environment. We address this issue in the context of Human-AI teaming by considering a setup where additional knowledge for the agent in the form of new candidate models arrives from a human operator in an online fashion. We introduce a new agent called DynamicHedgeAIXI that maintains an exact Bayesian mixture over dynamically changing sets of models via a time-adaptive prior constructed from a variant of the Hedge algorithm. The DynamicHedgeAIXI agent is the richest direct approximation of AIXI known to date and comes with good performance guarantees. Experimental results on epidemic control on contact networks validates the agent's practical utility.

It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are powerful general-purpose predictors and that the compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively. Finally, we show that the prediction-compression equivalence allows us to use any compressor (like gzip) to build a conditional generative model.

Meta-learning has emerged as a powerful approach to train neural networks to learn new tasks quickly from limited data by pre-training them on a broad set of tasks. But, what are the limits of meta-learning? In this work, we explore the potential of amortizing the most powerful universal predictor, namely Solomonoff Induction (SI), into neural networks via leveraging (memory-based) meta-learning to its limits. We use Universal Turing Machines (UTMs) to generate training data used to expose networks to a broad range of patterns. We provide theoretical analysis of the UTM data generation processes and meta-training protocols. We conduct comprehensive experiments with neural architectures (e.g. LSTMs, Transformers) and algorithmic data generators of varying complexity and universality. Our results suggest that UTM data is a valuable resource for meta-learning, and that it can be used to train neural networks capable of learning universal prediction strategies.

We propose a novel algorithmic framework for distributional reinforcement learning, based on learning finite-dimensional mean embeddings of return distributions. The framework reveals a wide variety of new algorithms for dynamic programming and temporal-difference algorithms that rely on the sketch Bellman operator, which updates mean embeddings with simple linear-algebraic computations. We provide asymptotic convergence theory, and examine the empirical performance of the algorithms on a suite of tabular tasks. Further, we show that this approach can be straightforwardly combined with deep reinforcement learning.

Inspired by recent progress in multi-agent Reinforcement Learning (RL), in this work we examine the collective intelligent behaviour of theoretical universal agents by introducing a weighted mixture operation. Given a weighted set of agents, their weighted mixture is a new agent whose expected total reward in any environment is the corresponding weighted average of the original agents’ expected total rewards in that environment. Thus, if RL agent intelligence is quantified in terms of performance across environments, the weighted mixture’s intelligence is the weighted average of the original agents’ intelligence. This operation enables various interesting new theorems that shed light on the geometry of RL agent intelligence, namely: results about symmetries, convex agent-sets, and local extrema. We also show that any RL agent intelligence measure based on average performance across environments, subject to certain weak technical conditions, is identical (up to a constant factor) to performance within a single environment dependent on said intelligence measure.

The Arcade Learning Environment (ALE) has become an essential benchmark for assessing the performance of reinforcement learning algorithms. However, the computational cost of generating results on the entire 57-game dataset limits ALE’s use and makes the reproducibility of many results infeasible. We propose a novel solution to this problem in the form of a principled methodology for selecting small but representative subsets of environments within a benchmark suite. We applied our method to identify a subset of five ALE games, we call Atari-5, which produces 57-game median score estimates within 10% of their true values. Extending the subset to 10-games recovers 80% of the variance for log-scores for all games within the 57-game set. We show this level of compression is possible due to a high degree of correlation between many of the games in ALE.

Memory-based meta-learning is a technique for approximating Bayes-optimal predictors. Under fairly general conditions, minimizing sequential prediction error, measured by the log loss, leads to implicit meta-learning. The goal of this work is to investigate how far this interpretation can be realized by current sequence prediction models and training regimes. The focus is on piecewise stationary sources with unobserved switching-points, which arguably capture an important characteristic of natural language and action-observation sequences in partially observable environments. We show that various types of memory-based neural models, including Transformers, LSTMs, and RNNs can learn to accurately approximate known Bayes-optimal algorithms and behave as if performing Bayesian inference over the latent switching-points and the latent parameters governing the data distribution within each segment.

Levin Tree Search (LTS) is a search algorithm that makes use of a policy (a probability distribution over actions) and comes with a theoretical guarantee on the number of expansions before reaching a goal node, depending on the quality of the policy. This guarantee can be used as a loss function, which we call the LTS loss, to optimize neural networks representing the policy (LTS+NN). In this work we show that the neural network can be substituted with parameterized context models originating from the online compression literature (LTS+CM). We show that the LTS loss is convex under this new model, which allows for using standard convex optimization tools, and obtain convergence guarantees to the optimal parameters in an online setting for a given set of solution trajectories --- guarantees that cannot be provided for neural networks. The new LTS+CM algorithm compares favorably against LTS+NN on several benchmarks: Sokoban (Boxoban), The Witness, and the 24-Sliding Tile puzzle (STP). The difference is particularly large on STP, where LTS+NN fails to solve most of the test instances while LTS+CM solves each test instance in a fraction of a second. Furthermore, we show that LTS+CM is able to learn a policy that solves the Rubik's cube in only a few hundred expansions, which considerably improves upon previous machine learning techniques.

Reinforcement Learning (RL) algorithms typically utilize learning and/or planning techniques to derive effective policies. The integration of both approaches has proven to be highly successful in addressing complex sequential decision-making challenges, as evidenced by algorithms such as AlphaZero and MuZero, which consolidate the planning process into a parametric search-policy. AIXI, the most potent theoretical universal agent, leverages planning through comprehensive search as its primary means to find an optimal policy. Here we define an alternative universal agent, which we call Self-AIXI, that on the contrary to AIXI, maximally exploits learning to obtain good policies. It does so by self-predicting its own stream of action data, which is generated, similarly to other TD(0) agents, by taking an action maximization step over the current on-policy (universal mixture-policy) Q-value estimates. We prove that Self-AIXI converges to AIXI, and inherits a series of properties like maximal Legg-Hutter intelligence and the self-optimizing property.

In imitation learning, imitators and demonstrators are policies for picking actions given past interactions with the environment. If we run an imitator, we probably want events to unfold similarly to the way they would have if the demonstrator had been acting the whole time. In general, one mistake during learning can lead to completely different events. In the special setting of environments that restart, existing work provides formal guidance in how to imitate so that events unfold similarly, but outside that setting, no formal guidance exists. We address a fully general setting, in which the (stochastic) environment and demonstrator never reset, not even for training purposes, and we allow our imitator to learn online from the demonstrator. Our new conservative Bayesian imitation learner underestimates the probabilities of each available action, and queries for more data with the remaining probability. Our main result: if an event would have been unlikely had the demonstrator acted the whole time, that event's likelihood can be bounded above when running the (initially totally ignorant) imitator instead. Meanwhile, queries to the demonstrator rapidly diminish in frequency. If any such event qualifies as "dangerous", our imitator would have the notable distinction of being relatively "safe".

The reinforcement learning (RL) framework formalizes the notion of learning with interactions. Many real-world problems have large state-spaces and/or action-spaces such as in Go, StarCraft, protein folding, and robotics or are non-Markovian, which cause significant challenges to RL algorithms. In this work we address the large action-space problem by sequentializing actions, which can reduce the action-space size significantly, even down to two actions at the expense of an increased planning horizon. We provide explicit and exact constructions and equivalence proofs for all quantities of interest for arbitrary history-based processes. In the case of MDPs, this could help RL algorithms that bootstrap. In this work we show how action-binarization in the non-MDP case can significantly improve Extreme State Aggregation (ESA) bounds. ESA allows casting any (non-MDP, non-ergodic, history-based) RL problem into a fixed-sized non-Markovian state-space with the help of a surrogate Markovian process. On the upside, ESA enjoys similar optimality guarantees as Markovian models do. But a downside is that the size of the aggregated state-space becomes exponential in the size of the action-space. In this work, we patch this issue by binarizing the action-space. We provide an upper bound on the number of states of this binarized ESA that is logarithmic in the original action-space size, a double-exponential improvement.

This paper presents a new family of backpropagation-free neural architectures, Gated Linear Networks (GLNs). What distinguishes GLNs from contemporary neural networks is the distributed and local nature of their credit assignment mechanism; each neuron directly predicts the target, forgoing the ability to learn feature representations in favor of rapid online learning. Individual neurons are able to model nonlinear functions via the use of data-dependent gating in conjunction with online convex optimization. We show that this architecture gives rise to universal learning capabilities in the limit, with effective model capacity increasing as a function of network size in a manner comparable with deep ReLU networks. Furthermore, we demonstrate that the GLN learning mechanism possesses extraordinary resilience to catastrophic forgetting, performing almost on par to an MLP with dropout and Elastic Weight Consolidation on standard benchmarks.

Credit assignment in reinforcement learning is the problem of measuring an action’s influence on future rewards. In particular, this requires separating skill from luck, i.e. disentangling the effect of an action on rewards from that of external factors and subsequent actions. To achieve this, we adapt the notion of counterfactuals from causality theory to a model-free RL setup. The key idea is to condition value functions on future events, by learning to extract relevant information from a trajectory. We formulate a family of policy gradient algorithms that use these future-conditional value functions as baselines or critics, and show that they are provably low variance. To avoid the potential bias from conditioning on future information, we constrain the hindsight information to not contain information about the agent’s actions. We demonstrate the efficacy and validity of our algorithm on a number of illustrative and challenging problems.

General intelligence, the ability to solve arbitrary solvable problems, is supposed by many to be artificially constructible. Narrow intelligence, the ability to solve a given particularly difficult problem, has seen impressive recent development. Notable examples include self-driving cars, Go engines, image classifiers, and translators. Artificial General Intelligence (AGI) presents dangers that narrow intelligence does not: if something smarter than us across every domain were indifferent to our concerns, it would be an existential threat to humanity, just as we threaten many species despite no ill will. Even the theory of how to maintain the alignment of an AGI's goals with our own has proven highly elusive. We present the first algorithm we are aware of for asymptotically unambitious AGI, where “unambitiousness” includes not seeking arbitrary power. Thus, we identify an exception to the Instrumental Convergence Thesis, which is roughly that by default, an AGI would seek power, including over us.

If we could define the set of all bad outcomes, we could hard-code an agent which avoids them; however, in sufficiently complex environments, this is infeasible. We do not know of any general-purpose approaches in the literature to avoiding novel failure modes. Motivated by this, we define an idealized Bayesian reinforcement learner which follows a policy that maximizes the worst-case expected reward over a set of world-models. We call this agent pessimistic, since it optimizes assuming the worst case. A scalar parameter tunes the agent’s pessimism by changing the size of the set of world-models taken into account. Our first main contribution is: given an assumption about the agent’s model class, a sufficiently pessimistic agent does not cause “unprecedented events” with probability $1-\delta$, whether or not designers know how to precisely specify those precedents they are concerned with. Since pessimism discourages exploration, at each timestep, the agent may defer to a mentor, who may be a human or some known-safe policy we would like to improve. Our other main contribution is that the agent’s policy’s value approaches at least that of the mentor, while the probability of deferring to the mentor goes to 0. In high-stakes environments, we might like advanced artificial agents to pursue goals cautiously, which is a non-trivial problem even if the agent were allowed arbitrary computing power; we present a formal solution.

The Lottery Ticket Hypothesis is a conjecture that every large neural network contains a subnetwork that, when trained in isolation, achieves comparable performance to the large network. An even stronger conjecture has been proven recently: Every sufficiently overparameterized network contains a subnetwork that, even without training, achieves comparable accuracy to the trained large network. This theorem, however, relies on a number of strong assumptions and guarantees a polynomial factor on the size of the large network compared to the target function. In this work, we remove the most limiting assumptions of this previous work while providing significantly tighter bounds: the overparameterized network only needs a logarithmic factor (in all variables but depth) number of neurons per weight of the target subnetwork.

We introduce a new and completely online contextual bandit algorithm called Gated Linear Contextual Bandits (GLCB). This algorithm is based on Gated Linear Networks (GLNs), a recently introduced deep learning architecture with properties well-suited to the online setting. Leveraging data-dependent gating properties of the GLN we are able to estimate prediction uncertainty with effectively zero algorithmic overhead. We empirically evaluate GLCB compared to 9 state-of-the-art algorithms that leverage deep neural networks, on a standard benchmark suite of discrete and continuous contextual bandit problems. GLCB obtains mean first-place despite being the only online method, and we further support these results with a theoretical study of its convergence properties.

Human intelligence is characterized not only by the capacity to learn complex skills, but the ability to rapidly adapt and acquire new skills within an ever-changing environment. In this work we study how the learning of modular solutions can allow for effective generalization to both unseen and potentially differently distributed data. Our main postulate is that the combination of task segmentation, modular learning and memory-based ensembling can give rise to generalization on an exponentially growing number of unseen tasks. We provide a concrete instantiation of this idea using a combination of: (1) the Forget-Me-Not Process, for task segmentation and memory based ensembling; and (2) Gated Linear Networks, which in contrast to contemporary deep learning techniques use a modular and local learning mechanism. We demonstrate that this system exhibits a number of desirable continual learning properties: robustness to catastrophic forgetting, no negative transfer and increasing levels of positive transfer as more tasks are seen. We show competitive performance against both offline and online methods on standard continual learning benchmarks.

Most real-world problems have huge state and/or action spaces. Therefore, a naive application of existing tabular solution methods is not tractable on such problems. Nonetheless, these solution methods are quite useful if an agent has access to a relatively small state-action space homomorphism of the true environment and near-optimal performance is guaranteed by the map. A plethora of research is focused on the case when the homomorphism is a Markovian representation of the underlying process. However, we show that nearoptimal performance is sometimes guaranteed even if the homomorphism is non-Markovian.

Reinforcement Learning agents are expected to eventually perform well. Typically, this takes the form of a guarantee about the asymptotic behavior of an algorithm given some assumptions about the environment. We present an algorithm for a policy whose value approaches the optimal value with probability 1 in all computable probabilistic environments, provided the agent has a bounded horizon. This is known as strong asymptotic optimality, and it was previously unknown whether it was possible for a policy to be strongly asymptotically optimal in the class of all computable probabilistic environments. Our agent, Inquisitive Reinforcement Learner (Inq), is more likely to explore the more it expects an exploratory action to reduce its uncertainty about which environment it is in, hence the term inquisitive. Exploring inquisitively is a strategy that can be applied generally; for more manageable environment classes, inquisitiveness is tractable. We conducted experiments in "grid-worlds" to compare the Inquisitive Reinforcement Learner to other weakly asymptotically optimal agents.

The convergence of many reinforcement learning (RL) algorithms with linear function approximation has been investigated extensively but most proofs assume that these methods converge to a unique solution. In this paper, we provide a complete characterization of non-uniqueness issues for a large class of reinforcement learning algorithms, simultaneously unifying many counter-examples to convergence in a theoretical framework. We achieve this by proving a new condition on features that can determine whether the convergence assumptions are valid or non-uniqueness holds. We consider a general class of RL methods, which we call natural algorithms, whose solutions are characterized as the fixed point of a projected Bellman equation. Our main result proves that natural algorithms converge to the correct solution if and only if all the value functions in the approximation space satisfy a certain shape. This implies that natural algorithms are, in general, inherently prone to converge to the wrong solution for most feature choices even if the value function can be represented exactly. Given our results, we show that state aggregation-based features are a safe choice for natural algorithms and also provide a condition for finding convergent algorithms under other feature constructions.

We study the problem of compressing piecewise i.i.d. sources, which models the practical application of jointly compressing multiple disparate data files. We establish that universal compression of piecewise i.i.d data is possible by modeling the data as a Markov process whose memory grows suitably with the size of the data using the Krichevsky-Trofimov (KT) estimator. The memory order is chosen large enough so that successful learning of the distribution of the each piece of the data from the corresponding contexts is possible for almost any realization of any piecewise i.i.d. data process. This is, a priori, a surprising result given that we are employing a stationary model to asymptotically optimally (model and) compress non-stationary data.

The development of Artificial General Intelligence (AGI) promises to be a major event. Along with its many potential benefits, it also raises serious safety concerns. The intention of this paper is to provide an easily accessible and up-to-date collection of references for the emerging field of AGI safety. A significant number of safety problems for AGI have been identified. We list these, and survey recent research on solving them. We also cover works on how best to think of AGI from the limited knowledge we have today, predictions for when AGI will first be created, and what will happen after its creation. Finally, we review the current public policy on AGI.

Temporal-difference (TD) learning is an attractive, computationally efficient framework for model- free reinforcement learning. Q-learning is one of the most widely used TD learning technique that enables an agent to learn the optimal action-value function, i.e. Q-value function. Contrary to its widespread use, Q-learning has only been proven to converge on Markov Decision Processes (MDPs) and Q-uniform abstractions of finite-state MDPs. On the other hand, most real-world problems are inherently non-Markovian: the full true state of the environment is not revealed by recent observations. In this paper, we investigate the behavior of Q-learning when applied to non-MDP and non-ergodic domains which may have infinitely many underlying states. We prove that the convergence guarantee of Q-learning can be extended to a class of such non-MDP problems, in particular, to some non-stationary domains. We show that state-uniformity of the optimal Q-value function is a necessary and sufficient condition for Q-learning to converge even in the case of infinitely many internal states.

This work investigates the convergence rate of learning the stationary distribution of finite-alphabet stationary ergodic sources using a binarized context-tree weighting approach. The binarized context-tree weighting (CTW) algorithm estimates the stationary distribution of a symbol as a product of conditional distributions of each component bit, which are determined in a sequential manner using the well known binary context-tree weighting method. We establish that CTW algorithm is a consistent estimator of the stationary distribution, and that the worst-case L1-prediction error between the CTW and frequency estimates using n source symbols each of which when binarized consists of k > 1 bits decays as Θ(√{2k[logn/n]})·.

Many state-of-the-art reinforcement learning (RL) algorithms typically assume that the environment is an ergodic Markov Decision Process (MDP). In contrast, the field of universal reinforcement learning (URL) is concerned with algorithms that make as few assumptions as possible about the environment. The universal Bayesian agent AIXI and a family of related URL algorithms have been developed in this setting. While numerous theoretical optimality results have been proven for these agents, there has been no empirical investigation of their behavior to date. We present a short and accessible survey of these URL algorithms under a unified notation and framework, along with results of some experiments that qualitatively illustrate some properties of the resulting policies, and their relative performance on partially-observable gridworld environments. We also present an open- source reference implementation of the algorithms which we hope will facilitate further understanding of, and experimentation with, these ideas.

We discuss some recent results on Thompson sampling for nonparametric reinforcement learning in countable classes of general stochastic environments. These environments can be non-Markovian, non-ergodic, and partially observable. We show that Thompson sampling learns the environment class in the sense that (1) asymptotically its value converges in mean to the optimal value and (2) given a recoverability assumption regret is sublinear. We conclude with a discussion about optimality in reinforcement learning.

We introduce a new count-based optimistic exploration algorithm for Reinforcement Learning (RL) that is feasible in environments with high-dimensional state-action spaces. The success of RL algorithms in these domains depends crucially on generalisation from limited training experience. Function approximation techniques enable RL agents to generalise in order to estimate the value of unvisited states, but at present few methods enable generalisation regarding uncertainty. This has prevented the combination of scalable RL algorithms with efficient exploration strategies that drive the agent to reduce its uncertainty. We present a new method for computing a generalised state visit-count, which allows the agent to estimate the uncertainty associated with any state. Our \phi-pseudocount achieves generalisation by exploiting same feature representation of the state space that is used for value function approximation. States that have less frequently observed features are deemed more uncertain. The \phi-Exploration-Bonus algorithm rewards the agent for exploring in feature space rather than in the untransformed state space. The method is simpler and less computationally expensive than some previous proposals, and achieves near state-of-the-art results on high-dimensional RL benchmarks.

This paper establishes for the first time the predictive performance of speed priors and their computational complexity. A speed prior is essentially a probability distribution that puts low probability on strings that are not efficiently computable. We propose a variant to the original speed prior (Schmidhuber, 2002), and show that our prior can predict sequences drawn from probability measures that are estimable in polynomial time. Our speed prior is computable in doubly-exponential time, but not in polynomial time. On a polynomial time computable sequence our speed prior is computable in exponential time. We show that Schmidhuber’s speed prior has better complexity under the same conditions; however, the question of its predictive properties remains open.

We present hierarchical rank pooling, a video sequence encoding method for activity recognition. It consists of a network of rank pooling functions which captures the dynamics of rich convolutional neural network features within a video sequence. By stacking non-linear feature functions and rank pooling over one another, we obtain a high capacity dynamic encoding mechanism, which is used for action recognition. We present a method for jointly learning the video representation and activity classifier parameters. Our method obtains state-of-the art results on three important activity recognition benchmarks: 76.7% on Hollywood2, 66.9% on HMDB51 and, 91.4% on UCF101.

A big open question of algorithmic information theory is the choice of the universal Turing machine (UTM). For Kolmogorov complexity and Solomonoff induction we have invariance theorems: the choice of the UTM changes bounds only by a constant. For the universally intelligent agent AIXI (Hutter, 2005) no invariance theorem is known. Our results are entirely negative: we discuss cases in which unlucky or adversarial choices of the UTM cause AIXI to misbehave drastically. We show that Legg-Hutter intelligence and thus balanced Pareto optimality is entirely subjective, and that every policy is Pareto optimal in the class of all computable environments. This undermines all existing optimality properties for AIXI. While it may still serve as a gold standard for AI, our results imply that AIXI is a \emphrelative theory, dependent on the choice of the UTM.

Function optimisation is a major challenge in computer science. The No Free Lunch theorems state that if all functions with the same histogram are assumed to be equally probable then no algorithm outperforms any other in expectation. We argue against the uniform assumption and suggest a universal prior exists for which there is a free lunch, but where no particular class of functions is favoured over another. We also prove upper and lower bounds on the size of the free lunch.

Can we measure the difficulty of an optimization problem? Although optimization plays a crucial role in modern science and technology, a formal framework that puts problems and solution algorithms into a broader context has not been established. This paper presents a conceptual approach which gives a positive answer to the question for a broad class of optimization problems. Adopting an information and computational perspective, the proposed framework builds upon Shannon and algorithmic information theories. As a starting point, a concrete model and definition of optimization problems is provided. Then, a formal definition of optimization difficulty is introduced which builds upon algorithmic information theory. Following an initial analysis, lower and upper bounds on optimization difficulty are established. One of the upper-bounds is closely related to Shannon information theory and black-box optimization. Finally, various computational issues and future research directions are discussed.

We extend the Q-learning algorithm from the Markov Decision Process setting to problems where observations are non-Markov and do not reveal the full state of the world i.e. to POMDPs. We do this in a natural manner by adding \ell_0 regularisation to the pathwise squared Q-learning objective function and then optimise this over both a choice of map from history to states and the resulting MDP parameters. The optimisation procedure involves a stochastic search over the map class nested with classical Q-learning of the parameters. This algorithm fits perfectly into the feature reinforcement learning framework, which chooses maps based on a cost criteria. The cost criterion used so far for feature reinforcement learning has been model-based and aimed at predicting future states and rewards. Instead we directly predict the return, which is what is needed for choosing optimal actions. Our Q-learning criteria also lends itself immediately to a function approximation setting where features are chosen based on the history. This algorithm is somewhat similar to the recent line of work on lasso temporal difference learning which aims at finding a small feature set with which one can perform policy evaluation. The distinction is that we aim directly for learning the Q-function of the optimal policy and we use \ell_0 instead of \ell_1 regularisation. We perform an experimental evaluation on classical benchmark domains and find improvement in convergence speed as well as in economy of the state representation. We also compare against MC-AIXI on the large Pocman domain and achieve competitive performance in average reward. We use less than half the CPU time and 36 times less memory. Overall, our algorithm hQL provides a better combination of computational, memory and data efficiency than existing algorithms in this setting.

Online estimation and modelling of i.i.d. data for shortsequences over large or complex “alphabets” is a ubiquitous (sub)problem in machine learning, information theory, data compression, statistical language processing, and document analysis. The Dirichlet-Multinomial distribution (also called Polya urn scheme) and extensions thereof are widely applied for online i.i.d. estimation. Good a-priori choices for the parameters in this regime are difficult to obtain though. I derive an optimal adaptive choice for the main parameter via tight, data-dependent redundancy bounds for a related model. The 1-line recommendation is to set the ’total mass’ = ’precision’ = ’concentration’ parameter to m/[2\ln\fracn+1m], where n is the (past) sample size and m the number of different symbols observed (so far). The resulting estimator is simple, online, fast,and experimental performance is superb.

We study the sample-complexity of reinforcement learning in a general setting without assuming ergodicity or finiteness of the environment. Instead, we define a topology on the space of environments and show that if an environment class is compact with respect to this topology then finite sample-complexity bounds are possible and give an algorithm achieving these bounds. We also show the existence of environment classes that are non-compact where finite sample-complexity bounds are not achievable. A lower bound is presented that matches the upper bound except for logarithmic factors.

Recent developments in reinforcement learning for non-Markovianproblems witness a surge in history-based methods, among which weare particularly interested in two frameworks, PhiMDP and MC-AIXI-CTW. PhiMDP attempts to reduce the general RL problem, where the environment's states and dynamics are both unknown, toan MDP, while MC-AIXI-CTW incrementally learns a mixture of contexttrees as its environment model. The main idea of PhiMDP is toconnect generic reinforcement learning with classical reinforcementlearning. The first implementation of PhiMDP relies on astochastic search procedure for finding a tree that minimizes acertain cost function. This does not guarantee finding theminimizing tree, or even a good one, given limited search time. As aconsequence it appears that the approach has difficulties with largedomains. MC-AIXI-CTW is attractive in that it can incrementally andanalytically compute the internal model through interactions withthe environment. Unfortunately, it is computationally demanding dueto requiring heavy planning simulations at every single time step.We devise a novel approach called CTMRL, which analytically andefficiently finds the cost-minimizing tree. Instead of thecontext-tree weighting method that MC-AIXI-CTW is based on, we usethe closely related context-tree maximizing algorithm that selectsjust one single tree. This approach falls under the PhiMDPframework, which allows the replacement of the costly planningcomponent of MC-AIXI-CTW with simple Q-Learning. Our empiricalinvestigation show that CTMRL finds policies of quality as good as MC-AIXI-CTW's on sixdomains including a challenging Pacman domain, but in an order ofmagnitude less time.

We describe an adaptive context tree weighting (ACTW) algorithm, as an extension to the standard context tree weighting (CTW) algorithm. Unlike the standard CTW algorithm, which weights all observations equally regardless of the depth, ACTW gives increasing weight to more recent observations, aiming to improve performance in cases where the input sequence is from a non-stationary distribution. Data compression results show ACTW variants improving over CTW on merged files from standard compression benchmark tests while never being significantly worse on any individual file.

This paper describes the Context Tree Switching technique, a modification of Context Tree Weighting for the prediction of binary, stationary, n-Markov sources. By modifying Context Tree Weighting's recursive weighting scheme, it is possible to mix over a strictly larger class of models without increasing the asymptotic time or space complexity of the original algorithm. We prove that this generalization preserves the desirable theoretical properties of Context Tree Weighting on stationary n-Markov sources, and show empirically that this new technique leads to consistent improvements over Context Tree Weighting as measured on the Calgary Corpus.

This paper introduces a principled approach for the design of a scalable general reinforcement learning agent. Our approach is based on a direct approximation of AIXI, a Bayesian optimality notion for general reinforcement learning agents. Previously, it has been unclear whether the theory of AIXI could motivate the design of practical algorithms. We answer this hitherto open question in the affirmative, by providing the first computationally feasible approximation to the AIXI agent. To develop our approximation, we introduce a new Monte-Carlo Tree Search algorithm along with an agent-specific extension to the Context Tree Weighting algorithm. Empirically, we present a set of encouraging results on a variety of stochastic and partially observable domains. We conclude by proposing a number of directions for future research.

This paper introduces a principled approach for the design of a scalable general reinforcement learning agent. This approach is based on a direct approximation of AIXI, a Bayesian optimality notion for general reinforcement learning agents. Previously, it has been unclear whether the theory of AIXI could motivate the design of practical algorithms. We answer this hitherto open question in the affirmative, by providing the first computationally feasible approximation to the AIXI agent. To develop our approximation, we introduce a Monte Carlo Tree Search algorithm along with an agent-specific extension of the Context Tree Weighting algorithm. Empirically, we present a set of encouraging results on a number of stochastic, unknown, and partially observable domains.

BACKGROUND: Cancer and other disorders are due to genomic lesions. SNP-microarrays are able to measure simultaneously both genotype and copy number (CN) at several Single Nucleotide Polymorphisms (SNPs) along the genome. CN is defined as the number of DNA copies, and the normal is two, since we have two copies of each chromosome. The genotype of a SNP is the status given by the nucleotides (alleles) which are present on the two copies of DNA. It is defined homozygous or heterozygous if the two alleles are the same or if they differ, respectively. Loss of heterozygosity (LOH) is the loss of the heterozygous status due to genomic events. Combining CN and LOH data, it is possible to better identify different types of genomic aberrations. For example, a long sequence of homozygous SNPs might be caused by either the physical loss of one copy or a uniparental disomy event (UPD), i.e. each SNP has two identical nucleotides both derived from only one parent. In this situation, the knowledge of the CN can help in distinguishing between these two events. RESULTS: To better identify genomic aberrations, we propose a method (called gBPCR) which infers the type of aberration occurred, taking into account all the possible influence in the microarray detection of the homozygosity status of the SNPs, resulting from an altered CN level. Namely, we model the distributions of the detected genotype, given a specific genomic alteration and we estimate the parameters involved on public reference datasets. The estimation is performed similarly to the modified Bayesian Piecewise Constant Regression, but with improved estimators for the detection of the breakpoints.Using artificial and real data, we evaluate the quality of the estimation of gBPCR and we also show that it outperforms other well-known methods for LOH estimation. CONCLUSIONS: We propose a method (gBPCR) for the estimation of both LOH and CN aberrations, improving their estimation by integrating both types of data and accounting for their relationships. Moreover, gBPCR performed very well in comparison with other methods for LOH estimation and the estimated CN lesions on real data have been validated with another technique.

The Minimum Description Length (MDL) principle selects the model that has the shortest code for data plus model. We show that for a countable class of models, MDL predictions are close to the true distribution in a strong sense. The result is completely general. No independence, ergodicity, stationarity, identifiability, or other assumption on the model class need to be made. More formally, we show that for any countable class of models, the distributions selected by MDL (or MAP) asymptotically predict (merge with) the true measure in the class in total variation distance. Implications for non-i.i.d. domains like time-series forecasting, discriminative learning, and reinforcement learning are discussed.

BACKGROUND: Some diseases, like tumors, can be related to chromosomal aberrations, leading to changes of DNA copy number. The copy number of an aberrant genome can be represented as a piecewise constant function, since it can exhibit regions of deletions or gains. Instead, in a healthy cell the copy number is two because we inherit one copy of each chromosome from each our parents. Bayesian Piecewise Constant Regression (BPCR) is a Bayesian regression method for data that are noisy observations of a piecewise constant function. The method estimates the unknown segment number, the endpoints of the segments and the value of the segment levels of the underlying piecewise constant function. The Bayesian Regression Curve (BRC) estimates the same data with a smoothing curve. However, in the original formulation, some estimators failed to properly determine the corresponding parameters. For example, the boundary estimator did not take into account the dependency among the boundaries and succeeded in estimating more than one breakpoint at the same position, losing segments. RESULTS: We derived an improved version of the BPCR (called mBPCR) and BRC, changing the segment number estimator and the boundary estimator to enhance the fitting procedure. We also proposed an alternative estimator of the variance of the segment levels, which is useful in case of data with high noise. Using artificial data, we compared the original and the modified version of BPCR and BRC with other regression methods, showing that our improved version of BPCR generally outperformed all the others. Similar results were also observed on real data. CONCLUSION: We propose an improved method for DNA copy number estimation, mBPCR, which performed very well compared to previously published algorithms. In particular, mBPCR was more powerful in the detection of the true position of the breakpoints and of small aberrations in very noisy data. Hence, from a biological point of view, our method can be very useful, for example, to find targets of genomic aberrations in clinical cancer samples.

Crucial to an evolutionary algorithm's performance is its selection scheme. We mathematically investigate the relation between polynomial rank and probabilistic tournament methods which are (respectively) generalisations of the popular linear ranking and tournament selection schemes. We show that every probabilistic tournament is equivalent to a unique polynomial rank scheme. In fact, we derived explicit operators for translating between these two types of selection. Of particular importance is that most linear and most practical quadratic rank schemes are probabilistic tournaments.

Suppose we are given two probability measures on the set of one-way infinite finite-alphabet sequences. Consider the question when one of the measures predicts the other, that is, when conditional probabilities converge (in a certain sense), if one of the measures is chosen to generate the sequence. This question may be considered a refinement of the problem of sequence prediction in its most general formulation: for a given class of probability measures, does there exist a measure which predicts all of the measures in the class? To address this problem, we find some conditions on local absolute continuity which are sufficient for prediction and generalize several different notions that are known to be sufficient for prediction. We also formulate some open questions to outline a direction for finding the conditions on classes of measures for which prediction is possible.

We derive an equation for temporal difference learning from statistical principles. Specifically, we start with the variational principle and then bootstrap to produce an updating rule for discounted state value estimates. The resulting equation is similar to the standard equation for temporal difference learning with eligibil- ity traces, so called TD(λ), however it lacks the parameter α that specifies the learning rate. In the place of this free parameter there is now an equation for the learning rate that is specific to each state transition. We experimentally test this new learning rule against TD(λ) and find that it offers superior performance in various settings. Finally, we make some preliminary investigations into how to extend our new temporal difference algorithm to reinforcement learning. To do this we combine our update equation with both Watkins’ Q(λ) and Sarsa(λ) and find that it again offers superior performance without a learning rate parameter.

In evolutionary algorithms, the fitness of a population increases with time by mutating and recombining individuals and by a biased selection of fitter individuals. The right selection pressure is critical in ensuring sufficient optimization progress on the one hand and in preserving genetic diversity to be able to escape from local optima on the other hand. Motivated by a universal similarity relation on the individuals, we propose a new selection scheme, which is uniform in the fitness values. It generates selection pressure toward sparsely populated fitness regions, not necessarily toward higher fitness, as is the case for all other selection schemes. We show analytically on a simple example that the new selection scheme can be much more effective than standard selection schemes. We also propose a new deletion scheme which achieves a similar result via deletion and show how such a scheme preserves genetic diversity more effectively than standard approaches. We compare the performance of the new schemes to tournament selection and random deletion on an artificial deceptive problem and a range of NP hard problems: traveling salesman, set covering, and satisfiability

A commonly experienced problem with population based optimisation methods is the gradual decline in population diversity that tends to occur over time. This can slow a system's progress or even halt it completely if the population converges on a local optimum from which it cannot escape. In this paper we present the Fitness Uniform Deletion Scheme (FUDS), a simple but somewhat unconventional approach to this problem. Under FUDS the deletion operation is modified to only delete those individuals which are "common" in the sense that there exist many other individuals of similar fitness in the population. This makes it impossible for the population to collapse to a collection of highly related individuals with similar fitness. Our experimental results on a range of optimisation problems confirm this, in particular for deceptive optimisation problems the performance is significantly more robust to variation in the selection intensity.

When applying aggregating strategies to Prediction with Expert Advice (PEA), the learning rate must be adaptively tuned. The natural choice of sqrt(complexity/current loss) renders the analysis of Weighted Majority (WM) derivatives quite complicated. In particular, for arbitrary weights there have been no results proven so far. The analysis of the alternative Follow the Perturbed Leader (FPL) algorithm from Kalai and Vempala (2003) based on Hannan's algorithm is easier. We derive loss bounds for adaptive learning rate and both finite expert classes with uniform weights and countable expert classes with arbitrary weights. For the former setup, our loss bounds match the best known results so far, while for the latter our results are new.

Minimum description length (MDL) is an important principle for induction and prediction, with strong relations to optimal Bayesian learning. This paper deals with learning processes which are independent and identically distributed (i.i.d.) by means of two-part MDL, where the underlying model class is countable. We consider the online learning framework, i.e., observations come in one by one, and the predictor is allowed to update its state of mind after each time step. We identify two ways of predicting by MDL for this setup, namely, a static and a dynamic one. (A third variant, hybrid MDL, will turn out inferior.) We will prove that under the only assumption that the data is generated by a distribution contained in the model class, the MDL predictions converge to the true values almost surely. This is accomplished by proving finite bounds on the quadratic, the Hellinger, and the Kullback-Leibler loss of the MDL learner, which are, however, exponentially worse than for Bayesian prediction. We demonstrate that these bounds are sharp, even for model classes containing only Bernoulli distributions. We show how these bounds imply regret bounds for arbitrary loss functions. Our results apply to a wide range of setups, namely, sequence prediction, pattern classification, regression, and universal induction in the sense of algorithmic information theory among others.

In evolutionary algorithms a critical parameter that must be tuned is that of selection pressure. If it is set too low then the rate of convergence towards the optimum is likely to be slow. Alternatively if the selection pressure is set too high the system is likely to become stuck in a local optimum due to a loss of diversity in the population. The recent fitness uniform selection scheme (FUSS) is a conceptually simple but somewhat radical approach to addressing this problem - rather than biasing the selection towards higher fitness, FUSS biases selection towards sparsely populated fitness levels. In this paper, we compare the relative performance of FUSS with the well known tournament selection scheme on a range of problems.

The probability of observing x/sub t/ at time t, given past observations x/sub 1/...x/sub t-1/ can be computed if the true generating distribution /spl mu/ of the sequences x/sub 1/x/sub 2/x/sub 3/... is known. If /spl mu/ is unknown, but known to belong to a class /spl Mscr/ one can base one's prediction on the Bayes mix /spl xi/ defined as a weighted sum of distributions /spl nu/ /spl isin/ /spl Mscr/. Various convergence results of the mixture posterior /spl xi//sub t/ to the true posterior /spl mu//sub t/ are presented. In particular, a new (elementary) derivation of the convergence /spl xi//sub t///spl mu//sub t/ /spl rarr/ 1 is provided, which additionally gives the rate of convergence. A general sequence predictor is allowed to choose an action y/sub t/ based on x/sub 1/...x/sub t-1/ and receives loss /spl lscr//sub x(t)y(t)/ if x/sub t/ is the next symbol of the sequence. No assumptions are made on the structure of /spl lscr/ (apart from being bounded) and /spl Mscr/. The Bayes-optimal prediction scheme /spl Lambda//sub /spl xi// based on mixture /spl xi/ and the Bayes-optimal informed prediction scheme /spl Lambda//sub /spl mu// are defined and the total loss L/sub /spl xi// of /spl Lambda//sub /spl xi// is bounded in terms of the total loss L/sub /spl mu// of /spl Lambda//sub /spl mu//. It is shown that L/sub /spl xi// is bounded for bounded L/sub /spl mu// and L/sub /spl xi///L/sub /spl mu// /spl rarr/ 1 for L/sub /spl mu// /spl rarr/ /spl infin/. Convergence of the instantaneous losses is also proven.

In evolutionary algorithms, the fitness of a population increases with time by mutating and recombining individuals and by a biased selection of more fit individuals. The right selection pressure is critical in ensuring sufficient optimization progress on the one hand and in preserving genetic diversity to be able to escape from local optima on the other. We propose a new selection scheme, which is uniform in the fitness values. It generates selection pressure towards sparsely populated fitness regions, not necessarily towards higher fitness, as is the case for all other selection schemes. We show that the new selection scheme can be more effective than standard selection schemes.

The mutual information of two random variables z and J with joint probabilities {7rij} is commonly used in learning Bayesian nets as well as in many other fields. The chances 7rij are usually estimated by the empirical sampling frequency nij In leading to a point es(cid:173) timate J(nij In) for the mutual information. To answer questions like "is J (nij In) consistent with zero?" or "what is the probability that the true mutual information is much larger than the point es(cid:173) timate?" one has to go beyond the point estimate. In the Bayesian framework one can answer these questions by utilizing a (second order) prior distribution p( 7r) comprising prior information about 7r. From the prior p(7r) one can compute the posterior p(7rln), from which the distribution p(Iln) of the mutual information can be cal(cid:173) culated. We derive reliable and quickly computable approximations for p(Iln). We concentrate on the mean, variance, skewness, and kurtosis, and non-informative priors. For the mean we also give an exact expression. Numerical issues and the range of validity are discussed.