Reinforcement learning, one of the most active research areas in artificial intelligence, is a computational approach to learning whereby an agent tries to maximize the total amount of reward it receives when interacting with a complex, uncertain environment. In Reinforcement Learning, Richard Sutton and Andrew Barto provide a clear and simple account of the key ideas and algorithms of reinforcement learning. Their discussion ranges from the history of the field's intellectual foundations to the most recent developments and applications. The only necessary mathematical background is familiarity with elementary concepts of probability. The book is divided into three parts. Part I defines the reinforcement learning problem in terms of Markov decision processes. Part II provides basic solution methods: dynamic programming, Monte Carlo methods, and temporal-difference learning. Part III presents a unified view of the solution methods and incorporates artificial neural networks, eligibility traces, and planning; the two final chapters present case studies and consider the future of reinforcement learning.

/pdf/reinforcement-learning-an-introduction-rzxgej9p17.pdf

Reinforcement Learning: An Introduction

https://www2.econ.iastate.edu/tesfatsi/DeepLearningInNeuralNetworksOverview.JSchmidhuber2015.pdf

Deep learning in neural networks

A long-standing goal of artificial intelligence is an algorithm that learns, tabula rasa, superhuman proficiency in challenging domains. Recently, AlphaGo became the first program to defeat a world champion in the game of Go. The tree search in AlphaGo evaluated positions and selected moves using deep neural networks. These neural networks were trained by supervised learning from human expert moves, and by reinforcement learning from self-play. Here we introduce an algorithm based solely on reinforcement learning, without human data, guidance or domain knowledge beyond game rules. AlphaGo becomes its own teacher: a neural network is trained to predict AlphaGo’s own move selections and also the winner of AlphaGo’s games. This neural network improves the strength of the tree search, resulting in higher quality move selection and stronger self-play in the next iteration. Starting tabula rasa, our new program AlphaGo Zero achieved superhuman performance, winning 100–0 against the previously published, champion-defeating AlphaGo. Starting from zero knowledge and without human data, AlphaGo Zero was able to teach itself to play Go and to develop novel strategies that provide new insights into the oldest of games. To beat world champions at the game of Go, the computer program AlphaGo has relied largely on supervised learning from millions of human expert moves. David Silver and colleagues have now produced a system called AlphaGo Zero, which is based purely on reinforcement learning and learns solely from self-play. Starting from random moves, it can reach superhuman level in just a couple of days of training and five million games of self-play, and can now beat all previous versions of AlphaGo. Because the machine independently discovers the same fundamental principles of the game that took humans millennia to conceptualize, the work suggests that such principles have some universal character, beyond human bias.

/pdf/mastering-the-game-of-go-without-human-knowledge-19y9mw638s.pdf

Mastering the game of Go without human knowledge

In this article, we describe a method for optimizing control policies, with guaranteed monotonic improvement. By making several approximations to the theoretically-justified scheme, we develop a practical algorithm, called Trust Region Policy Optimization (TRPO). This algorithm is effective for optimizing large nonlinear policies such as neural networks. Our experiments demonstrate its robust performance on a wide variety of tasks: learning simulated robotic swimming, hopping, and walking gaits; and playing Atari games using images of the screen as input. Despite its approximations that deviate from the theory, TRPO tends to give monotonic improvement, with little tuning of hyperparameters.

/pdf/trust-region-policy-optimization-3bqnki35vf.pdf

Trust Region Policy Optimization

We describe an iterative procedure for optimizing policies, with guaranteed monotonic improvement. By making several approximations to the theoretically-justified procedure, we develop a practical algorithm, called Trust Region Policy Optimization (TRPO). This algorithm is similar to natural policy gradient methods and is effective for optimizing large nonlinear policies such as neural networks. Our experiments demonstrate its robust performance on a wide variety of tasks: learning simulated robotic swimming, hopping, and walking gaits; and playing Atari games using images of the screen as input. Despite its approximations that deviate from the theory, TRPO tends to give monotonic improvement, with little tuning of hyperparameters.

We propose a new approach to reinforcement learning for control problems which combines value-function approximation with linear architectures and approximate policy iteration. This new approach is motivated by the least-squares temporal-difference learning algorithm (LSTD) for prediction problems, which is known for its efficient use of sample experiences compared to pure temporal-difference algorithms. Heretofore, LSTD has not had a straightforward application to control problems mainly because LSTD learns the state value function of a fixed policy which cannot be used for action selection and control without a model of the underlying process. Our new algorithm, least-squares policy iteration (LSPI), learns the state-action value function which allows for action selection without a model and for incremental policy improvement within a policy-iteration framework. LSPI is a model-free, off-policy method which can use efficiently (and reuse in each iteration) sample experiences collected in any manner. By separating the sample collection method, the choice of the linear approximation architecture, and the solution method, LSPI allows for focused attention on the distinct elements that contribute to practical reinforcement learning. LSPI is tested on the simple task of balancing an inverted pendulum and the harder task of balancing and riding a bicycle to a target location. In both cases, LSPI learns to control the pendulum or the bicycle by merely observing a relatively small number of trials where actions are selected randomly. LSPI is also compared against Q-learning (both with and without experience replay) using the same value function architecture. While LSPI achieves good performance fairly consistently on the difficult bicycle task, Q-learning variants were rarely able to balance for more than a small fraction of the time needed to reach the target location.

/pdf/least-squares-policy-iteration-d5xzu8eylp.pdf

Least-squares policy iteration

We present several new algorithms for multiagent reinforcement learning. A common feature of these algorithms is a parameterized, structured representation of a policy or value function. This structure is leveraged in an approach we call coordinated reinforcement learning, by which agents coordinate both their action selection activities and their parameter updates. Within the limits of our parametric representations, the agents will determine a jointly optimal action without explicitly considering every possible action in their exponentially large joint action space. Our methods differ from many previous reinforcement learning approaches to multiagent coordination in that structured communication and coordination between agents appears at the core of both the learning algorithm and the execution architecture.

/pdf/coordinated-reinforcement-learning-32ssfzs6j3.pdf

Coordinated Reinforcement Learning

Recently, auction methods have been investigated as effective, decentralized methods for multi-robot coordination. Experimental research has shown great potential, but has not been complemented yet by theoretical analysis. In this paper we contribute a theoretical analysis of the performance of auction methods for multi-robot routing. We suggest a generic framework for auction-based multi-robot routing and analyze a variety of bidding rules for different team objectives. This is the first time that auction methods are shown to offer theoretical guarantees for such a variety of bidding rules and team objectives.

https://www2.isye.gatech.edu/~pinar/robotics2005.pdf

Auction-Based Multi-Robot Routing

We consider the problem of allocating a number of exploration tasks to a team of mobile robots. Each task consists of a target location that needs to be visited by a robot. The objective of the allocation is to minimize the total cost, that is, the sum of the travel costs of all robots for visiting all targets. We show that finding an optimal allocation is an NP-hard problem, even in known environments. The main contribution of this paper is PRIM ALLOCATION, a simple and fast approximate algorithm for allocating targets to robots which provably computes allocations whose total cost is at most twice as large as the optimal total cost. We then cast PRIM ALLOCATION in terms of a multi-round single-item auction where robots bid on targets, which allow for a decentralized implementation. To the best of our knowledge, PRIM ALLOCATION is the first auction-based allocation algorithm that provides a guarantee on the quality of its allocations. Our experimental results in a multi-robot simulator demonstrate that PRIM ALLOCATION is fast and results in close-to-optimal allocations despite its simplicity and decentralized nature. In particular, it needs an order of magnitude fewer bids than a computationally intensive allocation algorithm based on combinatorial auctions, yet its allocations are at least as good.

/pdf/simple-auctions-with-performance-guarantees-for-multi-robot-alk6w6oh3s.pdf

Simple auctions with performance guarantees for multi-robot task allocation

The basic tools of machine learning appear in the inner loop of most reinforcement learning algorithms, typically in the form of Monte Carlo methods or function approximation techniques. To a large extent, however, current reinforcement learning algorithms draw upon machine learning techniques that are at least ten years old and, with a few exceptions, very little has been done to exploit recent advances in classification learning for the purposes of reinforcement learning. We use a variant of approximate policy iteration based on rollouts that allows us to use a pure classification learner, such as a support vector machine (SVM), in the inner loop of the algorithm. We argue that the use of SVMs, particularly in combination with the kernel trick, can make it easier to apply reinforcement learning as an "out-of-the-box" technique, without extensive feature engineering. Our approach opens the door to modern classification methods, but does not preclude the use of classical methods. We present experimental results in the pendulum balancing and bicycle riding domains using both SVMs and neural networks for classifiers.

/pdf/reinforcement-learning-as-classification-leveraging-modern-19gmrt8dpn.pdf

Michail G. Lagoudakis

Papers

Least-squares policy iteration

Coordinated Reinforcement Learning

Auction-Based Multi-Robot Routing

Simple auctions with performance guarantees for multi-robot task allocation

Reinforcement learning as classification: leveraging modern classifiers