One problem in the application of reinforcement learning to real-world problems is the curse of dimensionality on the action space. Macro actions, a sequence of primitive actions, have been studied to diminish the dimensionality of the action space with regard to the time axis. However, previous studies relied on humans defining macro actions or assumed macro actions as repetitions of the same primitive actions. We present Factorized Macro Action Reinforcement Learning (FaMARL) which autonomously learns disentangled factor representation of a sequence of actions to generate macro actions that can be directly applied to general reinforcement learning algorithms. FaMARL exhibits higher scores than other reinforcement learning algorithms on environments that require an extensive amount of search.
Numerous past works have tackled the problem of task-driven navigation. But, how to effectively explore a new environment to enable a variety of down-stream tasks has received much less attention. In this work, we study how agents can autonomously explore realistic and complex 3D environments without the context of task-rewards. We propose a learning-based approach and investigate different policy architectures, reward functions, and training paradigms. We find that the use of policies with spatial memory that are bootstrapped with imitation learning and finally finetuned with coverage rewards derived purely from on-board sensors can be effective at exploring novel environments. We show that our learned exploration policies can explore better than classical approaches based on geometry alone and generic learning-based exploration techniques. Finally, we also show how such task-agnostic exploration can be used for down-stream tasks. Code and Videos are available at: https://sites.google.com/view/exploration-for-nav.
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.
Value functions derived from Markov decision processes arise as a central component of algorithms as well as performance metrics in many statistics and engineering applications of machine learning techniques. Computation of the solution to the associated Bellman equations is challenging in most practical cases of interest. A popular class of approximation techniques, known as Temporal Difference (TD) learning algorithms, are an important sub-class of general reinforcement learning methods. The algorithms introduced in this paper are intended to resolve two well-known difficulties of TD-learning approaches: Their slow convergence due to very high variance, and the fact that, for the problem of computing the relative value function, consistent algorithms exist only in special cases. First we show that the gradients of these value functions admit a representation that lends itself to algorithm design. Based on this result, a new class of differential TD-learning algorithms is introduced. For Markovian models on Euclidean space with smooth dynamics, the algorithms are shown to be consistent under general conditions. Numerical results show dramatic variance reduction when compared to standard methods.
Machine learning applications in medical imaging are frequently limited by the lack of quality labeled data. In this paper, we explore the self training method, a form of semi-supervised learning, to address the labeling burden. By integrating reinforcement learning, we were able to expand the application of self training to complex segmentation networks without any further human annotation. The proposed approach, reinforced self training (ReST), fine tunes a semantic segmentation networks by introducing a policy network that learns to generate pseudolabels. We incorporate an expert demonstration network, based on inverse reinforcement learning, to enhance clinical validity and convergence of the policy network. The model was tested on a pulmonary nodule segmentation task in chest X-rays and achieved the performance of a standard U-Net while using only 50% of the labeled data, by exploiting unlabeled data. When the same number of labeled data was used, a moderate to significant cross validation accuracy improvement was achieved depending on the absolute number of labels used.
Developing visual perception models for active agents and sensorimotor control are cumbersome to be done in the physical world, as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning-in-simulation which consequently casts a question on whether the results transfer to real-world. In this paper, we are concerned with the problem of developing real-world perception for active agents, propose Gibson Virtual Environment for this purpose, and showcase sample perceptual tasks learned therein. Gibson is based on virtualizing real spaces, rather than using artificially designed ones, and currently includes over 1400 floor spaces from 572 full buildings. The main characteristics of Gibson are: I. being from the real-world and reflecting its semantic complexity, II. having an internal synthesis mechanism, "Goggles", enabling deploying the trained models in real-world without needing further domain adaptation, III. embodiment of agents and making them subject to constraints of physics and space.
We introduce a dynamic neural algorithm called Dynamic Neural (DN) SARSA(\lambda) for learning a behavioral sequence from delayed reward. DN-SARSA(\lambda) combines Dynamic Field Theory models of behavioral sequence representation, classical reinforcement learning, and a computational neuroscience model of working memory, called Item and Order working memory, which serves as an eligibility trace. DN-SARSA(\lambda) is implemented on both a simulated and real robot that must learn a specific rewarding sequence of elementary behaviors from exploration. Results show DN-SARSA(\lambda) performs on the level of the discrete SARSA(\lambda), validating the feasibility of general reinforcement learning without compromising neural dynamics.
We present a new algorithm for general reinforcement learning where the true environment is known to belong to a finite class of N arbitrary models. The algorithm is shown to be near-optimal for all but O(N log^2 N) time-steps with high probability. Infinite classes are also considered where we show that compactness is a key criterion for determining the existence of uniform sample-complexity bounds. A matching lower bound is given for the finite case.
We propose a general framework for sequential and dynamic acquisition of useful information in order to solve a particular task. While our goal could in principle be tackled by general reinforcement learning, our particular setting is constrained enough to allow more efficient algorithms. In this paper, we work under the Learning to Search framework and show how to formulate the goal of finding a dynamic information acquisition policy in that framework. We apply our formulation on two tasks, sentiment analysis and image recognition, and show that the learned policies exhibit good statistical performance. As an emergent byproduct, the learned policies show a tendency to focus on the most prominent parts of each instance and give harder instances more attention without explicitly being trained to do so.
Any agent that is part of the environment it interacts with and has versatile actuators (such as arms and fingers), will in principle have the ability to self-modify -- for example by changing its own source code. As we continue to create more and more intelligent agents, chances increase that they will learn about this ability. The question is: will they want to use it? For example, highly intelligent systems may find ways to change their goals to something more easily achievable, thereby `escaping' the control of their designers. In an important paper, Omohundro (2008) argued that goal preservation is a fundamental drive of any intelligent system, since a goal is more likely to be achieved if future versions of the agent strive towards the same goal. In this paper, we formalise this argument in general reinforcement learning, and explore situations where it fails. Our conclusion is that the self-modification possibility is harmless if and only if the value function of the agent anticipates the consequences of self-modifications and use the current utility function when evaluating the future.
Reinforcement learning (RL) problems are often phrased in terms of Markov decision processes (MDPs). In this thesis we go beyond MDPs and consider RL in environments that are non-Markovian, non-ergodic and only partially observable. Our focus is not on practical algorithms, but rather on the fundamental underlying problems: How do we balance exploration and exploitation? How do we explore optimally? When is an agent optimal? We follow the nonparametric realizable paradigm. We establish negative results on Bayesian RL agents, in particular AIXI. We show that unlucky or adversarial choices of the prior cause the agent to misbehave drastically. Therefore Legg-Hutter intelligence and balanced Pareto optimality, which depend crucially on the choice of the prior, are entirely subjective. Moreover, in the class of all computable environments every policy is Pareto optimal. This undermines all existing optimality properties for AIXI. However, there are Bayesian approaches to general RL that satisfy objective optimality guarantees: We prove that Thompson sampling is asymptotically optimal in stochastic environments in the sense that its value converges to the value of the optimal policy. We connect asymptotic optimality to regret given a recoverability assumption on the environment that allows the agent to recover from mistakes. Hence Thompson sampling achieves sublinear regret in these environments. Our results culminate in a formal solution to the grain of truth problem: A Bayesian agent acting in a multi-agent environment learns to predict the other agents' policies if its prior assigns positive probability to them (the prior contains a grain of truth). We construct a large but limit computable class containing a grain of truth and show that agents based on Thompson sampling over this class converge to play Nash equilibria in arbitrary unknown computable multi-agent environments.
In recent years, work has been done to develop the theory of General Reinforcement Learning (GRL). However, there are few examples demonstrating these results in a concrete way. In particular, there are no examples demonstrating the known results regarding gener- alised discounting. We have added to the GRL simulation platform AIXIjs the functionality to assign an agent arbitrary discount functions, and an environment which can be used to determine the effect of discounting on an agent's policy. Using this, we investigate how geometric, hyperbolic and power discounting affect an informed agent in a simple MDP. We experimentally reproduce a number of theoretical results, and discuss some related subtleties. It was found that the agent's behaviour followed what is expected theoretically, assuming appropriate parameters were chosen for the Monte-Carlo Tree Search (MCTS) planning algorithm.
Reinforcement learning is a general and powerful framework with which to study and implement artificial intelligence. Recent advances in deep learning have enabled RL algorithms to achieve impressive performance in restricted domains such as playing Atari video games (Mnih et al., 2015) and, recently, the board game Go (Silver et al., 2016). However, we are still far from constructing a generally intelligent agent. Many of the obstacles and open questions are conceptual: What does it mean to be intelligent? How does one explore and learn optimally in general, unknown environments? What, in fact, does it mean to be optimal in the general sense? The universal Bayesian agent AIXI (Hutter, 2005) is a model of a maximally intelligent agent, and plays a central role in the sub-field of general reinforcement learning (GRL). Recently, AIXI has been shown to be flawed in important ways; it doesn't explore enough to be asymptotically optimal (Orseau, 2010), and it can perform poorly with certain priors (Leike and Hutter, 2015). Several variants of AIXI have been proposed to attempt to address these shortfalls: among them are entropy-seeking agents (Orseau, 2011), knowledge-seeking agents (Orseau et al., 2013), Bayes with bursts of exploration (Lattimore, 2013), MDL agents (Leike, 2016a), Thompson sampling (Leike et al., 2016), and optimism (Sunehag and Hutter, 2015). We present AIXIjs, a JavaScript implementation of these GRL agents. This implementation is accompanied by a framework for running experiments against various environments, similar to OpenAI Gym (Brockman et al., 2016), and a suite of interactive demos that explore different properties of the agents, similar to REINFORCEjs (Karpathy, 2015). We use AIXIjs to present numerous experiments illustrating fundamental properties of, and differences between, these agents.
This paper introduces Dex, a reinforcement learning environment toolkit specialized for training and evaluation of continual learning methods as well as general reinforcement learning problems. We also present the novel continual learning method of incremental learning, where a challenging environment is solved using optimal weight initialization learned from first solving a similar easier environment. We show that incremental learning can produce vastly superior results than standard methods by providing a strong baseline method across ten Dex environments. We finally develop a saliency method for qualitative analysis of reinforcement learning, which shows the impact incremental learning has on network attention.
Speech recognition systems have achieved high recognition performance for several tasks. However, the performance of such systems is dependent on the tremendously costly development work of preparing vast amounts of task-matched transcribed speech data for supervised training. The key problem here is the cost of transcribing speech data. The cost is repeatedly required to support new languages and new tasks. Assuming broad network services for transcribing speech data for many users, a system would become more self-sufficient and more useful if it possessed the ability to learn from very light feedback from the users without annoying them. In this paper, we propose a general reinforcement learning framework for speech recognition systems based on the policy gradient method. As a particular instance of the framework, we also propose a hypothesis selection-based reinforcement learning method. The proposed framework provides a new view for several existing training and adaptation methods. The experimental results show that the proposed method improves the recognition performance compared to unsupervised adaptation.
A major bottleneck for developing general reinforcement learning agents is determining rewards that will yield desirable behaviors under various circumstances. We introduce a general mechanism for automatically specifying meaningful behaviors from raw pixels. In particular, we train a generative adversarial network to produce short sub-goals represented through motion templates. We demonstrate that this approach generates visually meaningful behaviors in unknown environments with novel agents and describe how these motions can be used to train reinforcement learning agents.
Discrete-action algorithms have been central to numerous recent successes of deep reinforcement learning. However, applying these algorithms to high-dimensional action tasks requires tackling the combinatorial increase of the number of possible actions with the number of action dimensions. This problem is further exacerbated for continuous-action tasks that require fine control of actions via discretization. In this paper, we propose a novel neural architecture featuring a shared decision module followed by several network branches, one for each action dimension. This approach achieves a linear increase of the number of network outputs with the number of degrees of freedom by allowing a level of independence for each individual action dimension. To illustrate the approach, we present a novel agent, called Branching Dueling Q-Network (BDQ), as a branching variant of the Dueling Double Deep Q-Network (Dueling DDQN). We evaluate the performance of our agent on a set of challenging continuous control tasks. The empirical results show that the proposed agent scales gracefully to environments with increasing action dimensionality and indicate the significance of the shared decision module in coordination of the distributed action branches. Furthermore, we show that the proposed agent performs competitively against a state-of-the-art continuous control algorithm, Deep Deterministic Policy Gradient (DDPG).
The game of chess is the most widely-studied domain in the history of artificial intelligence. The strongest programs are based on a combination of sophisticated search techniques, domain-specific adaptations, and handcrafted evaluation functions that have been refined by human experts over several decades. In contrast, the AlphaGo Zero program recently achieved superhuman performance in the game of Go, by tabula rasa reinforcement learning from games of self-play. In this paper, we generalise this approach into a single AlphaZero algorithm that can achieve, tabula rasa, superhuman performance in many challenging domains. Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case.
In reinforcement learning, it is common to let an agent interact for a fixed amount of time with its environment before resetting it and repeating the process in a series of episodes. The task that the agent has to learn can either be to maximize its performance over (i) that fixed period, or (ii) an indefinite period where time limits are only used during training to diversify experience. In this paper, we provide a formal account for how time limits could effectively be handled in each of the two cases and explain why not doing so can cause state aliasing and invalidation of experience replay, leading to suboptimal policies and training instability. In case (i), we argue that the terminations due to time limits are in fact part of the environment, and thus a notion of the remaining time should be included as part of the agent's input to avoid violation of the Markov property. In case (ii), the time limits are not part of the environment and are only used to facilitate learning. We argue that this insight should be incorporated by bootstrapping from the value of the state at the end of each partial episode. For both cases, we illustrate empirically the significance of our considerations in improving the performance and stability of existing reinforcement learning algorithms, showing state-of-the-art results on several control tasks.