While exploring to find better solutions, an agent performing online reinforcement learning (RL) can perform worse than is acceptable. In some cases, exploration might have unsafe, or even catastrophic, results, often modeled in terms of reaching'failure' states of the agent's environment. This paper presents a method that uses domain knowledge to reduce the number of failures during exploration. This method formulates the set of actions from which the RL agent composes a control policy to ensure that exploration is conducted in a policy space that excludes most of the unacceptable policies. The resulting action set has a more abstract relationship to the task being solved than is common in many applications of RL. Although the cost of this added safety is that learning may result in a suboptimal solution, we argue that this is an appropriate tradeoff in many problems. We illustrate this method in the domain of motion planning. "'This work was done while the first author was finishing his Ph.D in computer science at the University of Massachusetts, Amherst.

Survival is enhanced by an ability to predict the availability of food, the likelihood of predators, and the presence of mates. We present a concrete model that uses diffuse neurotransmitter systems to implement a predictive version of a Hebb learning rule embedded in a neural architecture based on anatomical and physiological studies on bees. The model captured the strategies seen in the behavior of bees and a number of other animals when foraging in an uncertain environment. The predictive model suggests a unified way in which neuromodulatory influences can be used to bias actions and control synaptic plasticity. Successful predictions enhance adaptive behavior by allowing organisms to prepare for future actions, rewards, or punishments. Moreover, it is possible to improve upon behavioral choices if the consequences of executing different actions can be reliably predicted. Although classical and instrumental conditioning results from the psychological literature [1] demonstrate that the vertebrate brain is capable of reliable prediction, how these predictions are computed in brains is not yet known. The brains of vertebrates and invertebrates possess small nuclei which project axons throughout large expanses of target tissue and deliver various neurotransmitters such as dopamine, norepinephrine, and acetylcholine [4]. The activity in these systems may report on reinforcing stimuli in the world or may reflect an expectation of future reward [5, 6,7,8].

The goal of this workshop was to discuss two major issues: efficient exploration of a learner's state space, and learning in continuous domains. The common themes that emerged in presentations and in discussion were the importance of choosing one's domain assumptions carefully, mixing controllers/strategies, avoidance of catastrophic failure, new approaches with difficulties with reinforcement learning, and the importance of task transfer. He suggested that neither "fewer assumptions are better" nor "more assumptions are better" is a tenable position, and that we should strive to find and use standard sets of assumptions. With no such commonality, comparison of techniques and results is meaningless. Under Moore's guidance, the group discussed the possibility of designing an algorithm which used a number of well-chosen assumption sets and switched between them according to their empirical validity.

Furthermore, we have verified that weights learned from 9x9 Go offer a suitable basis for further training on the full-size (19x19) board.

Parti-game is a new algorithm for learning from delayed rewards in high dimensional real-valued state-spaces. In high dimensions it is essential that learning does not explore or plan over state space uniformly. Part i-game maintains a decision-tree partitioning of state-space and applies game-theory and computational geometry techniques to efficiently and reactively concentrate high resolution only on critical areas. Many simulated problems have been tested, ranging from 2-dimensional to 9-dimensional state-spaces, including mazes, path planning, nonlinear dynamics, and uncurling snake robots in restricted spaces. In all cases, a good solution is found in less than twenty trials and a few minutes. 1 REINFORCEMENT LEARNING Reinforcement learning [Samuel, 1959, Sutton, 1984, Watkins, 1989, Barto et al., 1991] is a promising method for control systems to program and improve themselves.

Increasing attention has recently been paid to algorithms based on dynamic programming (DP) due to the suitability of DP for learning problems involving control. In stochastic environments where the system being controlled is only incompletely known, however, a unifying theoretical account of these methods has been missing. In this paper we relate DPbased learning algorithms to the powerful techniques of stochastic approximation via a new convergence theorem, enabling us to establish a class of convergent algorithms to which both TD("\) and Q-Iearning belong. 1 INTRODUCTION Learning to predict the future and to find an optimal way of controlling it are the basic goals of learning systems that interact with their environment. A variety of algorithms are currently being studied for the purposes of prediction and control in incompletely specified, stochastic environments. Here we consider learning algorithms defined in Markov environments. There are actions or controls (u) available for the learner that affect both the state transition probabilities, and the probability distribution for the immediate, state dependent costs (Ci(u)) incurred by the learner.

Reinforcement Learning methods based on approximating dynamic programming (DP) are receiving increased attention due to their utility in forming reactive control policies for systems embedded in dynamic environments. Environments are usually modeled as controlled Markov processes, but when the environment model is not known a priori, adaptive methods are necessary. Adaptive control methods are often classified as being direct or indirect. Direct methods directly adapt the control policy from experience, whereas indirect methods adapt a model of the controlled process and compute control policies based on the latest model. Our focus is on indirect adaptive DPbased methods in this paper. We present a convergence result for indirect adaptive asynchronous value iteration algorithms for the case in which a lookup table is used to store the value function. Our result implies convergence of several existing reinforcement learning algorithms such as adaptive real-time dynamic programming (ARTDP) (Barto, Bradtke, & Singh, 1993) and prioritized sweeping (Moore & Atkeson, 1993). Although the emphasis of researchers studying DPbased reinforcement learning has been on direct adaptive methods such as Q-Learning (Watkins, 1989) and methods using TD algorithms (Sutton, 1988), it is not clear that these direct methods are preferable in practice to indirect methods such as those analyzed in this paper.

We describe the relationship between certain reinforcement learning (RL) methods based on dynamic programming (DP) and a class of unorthodox Monte Carlo methods for solving systems of linear equations proposed in the 1950's. These methods recast the solution of the linear system as the expected value of a statistic suitably defined over sample paths of a Markov chain. The significance of our observations lies in arguments (Curtiss, 1954) that these Monte Carlo methods scale better with respect to state-space size than do standard, iterative techniques for solving systems of linear equations. This analysis also establishes convergence rate estimates. Because methods used in RL systems for approximating the evaluation function of a fixed control policy also approximate solutions to systems of linear equations, the connection to these Monte Carlo methods establishes that algorithms very similar to TD algorithms (Sutton, 1988) are asymptotically more efficient in a precise sense than other methods for evaluating policies. Further, all DPbased RL methods have some of the properties of these Monte Carlo algorithms, which suggests that although RL is often perceived to be slow, for sufficiently large problems, it may in fact be more efficient than other known classes of methods capable of producing the same results.

The field of reinforcement learning has grown dramatically over the past several years, but with the exception of backgammon [8, 2], has had few successful applications to large-scale, practical tasks. This paper demonstrates that the practical task of routing packets through a communication network is a natural application for reinforcement learning algorithms.

While exploring to find better solutions, an agent performing online reinforcement learning (RL) can perform worse than is acceptable. In some cases, exploration might have unsafe, or even catastrophic, results, often modeled in terms of reaching'failure' states of the agent's environment. This paper presents a method that uses domain knowledge to reduce the number of failures during exploration. This method formulates the set of actions from which the RL agent composes a control policy to ensure that exploration is conducted in a policy space that excludes most of the unacceptable policies. The resulting action set has a more abstract relationship to the task being solved than is common in many applications of RL. Although the cost of this added safety is that learning may result in a suboptimal solution, we argue that this is an appropriate tradeoff in many problems. We illustrate this method in the domain of motion planning. "'This work was done while the first author was finishing his Ph.D in computer science at the University of Massachusetts, Amherst.