This paper discusses theoretical and experimental aspects of gradient-based approaches to the direct optimization of policy performance in controlled POMDPs. We introduce GPOMDP, a REINFORCE-like algorithm for estimating an approximation to the gradient of the average reward as a function of the parameters of a stochastic policy. The algorithm's chief advantages are that it requires only a single sample path of the underlying Markov chain, it uses only one free parameter 2 [0; 1), which has a natural interpretation in terms of bias-variance trade-off, and it requires no knowledge of the underlying state. We prove convergence of GPOMDP and show how the gradient estimates produced by GPOMDP can be used in a conjugate-gradient procedure to find local optima of the average reward. 1. Introduction "Reinforcement learning" is used to describe the general problem of training an agent to choose its actions so as to increase its long-term average reward. The structure of th...

In [2] we introduced  ¢¡¤£¦¥¨§¦¡, an algorithm for computing arbitrarily accurate approximations to the performance gradient of parameterized partially observable Markov decision processes ( ¡©£¦¥¨§¦ ¡ s). The algorithm’s chief advantages are that it requires only a single sample path of the underlying Markov chain, it uses only one ���� � ������ � free parameter which has a natural interpretation in terms of bias-variance trade-off, and it requires no knowledge of the underlying state. In addition, the algorithm can be applied to infinite state, control and observation spaces.

In this paper, we present algorithms that perform gradient ascent of the average reward in a partially observable Markov decision process (POMDP). These algorithms are based on GPOMDP, an algorithm introduced in a companion paper (Baxter & Bartlett, 2001), which computes biased estimates of the performance gradient in POMDPs. The algorithm's chief advantages are that it uses only one free parameter 2 [0; 1), which has a natural interpretation in terms of bias-variance trade-off, it requires no knowledge of the underlying state, and it can be applied to infinite state, control and observation spaces. We show how the gradient estimates produced by GPOMDP can be used to perform gradient ascent, both with a traditional stochastic-gradient algorithm, and with an algorithm based on conjugate-gradients that utilizes gradient information to bracket maxima in line searches. Experimental results are presented illustrating both the theoretical results of Baxter and Bartlett (2001) on a toy problem, and practical aspects of the algorithms on a number of more realistic problems. 1.

The ease or difficulty in solving a problem strongly depends on the way it is represented. For example, consider the task of multiplying the numbers 12 and 24. Now imagine multiplying XII and XXIV. Both tasks can be solved, but it is clearly more difficult to use the Roman numeral representations of twelve and twenty-four. Humans excel at finding appropriate representations for solving complex problems. This is not true for artificial systems, which have largely relied on humans to provide appropriate representations. The ability to autonomously construct useful representations and to efficiently exploit them is an important challenge for artificial intelligence. This dissertation builds on a recently introduced graph-based approach to learning representations for sequential decision-making problems modeled as Markov decision processes (MDPs). Representations, or basis functions, forMDPs are abstractions of the problem’s state space and are used to approximate value functions, which quantify the expected long-term utility obtained by following a policy. The graph-based approach generates basis functions capturing the structure of the environment. Handling large environments requires efficiently constructing and utilizing these functions. We address two issues with this approach: (1) scaling basis construction and value function approximation to large graphs/data sets, and (2) tailoring the approximation to a specific policy’s value function. We introduce two algorithms for computing basis functions from large graphs. Both algorithms work by decomposing the basis construction problem into smaller, more manageable subproblems. One method determines the subproblems by enforcing block structure, or groupings of states. The other method uses recursion to solve subproblems which are then used for approximating the original problem. Both algorithms result in a set of basis functions from which we employ basis selection algorithms. The selection algorithms represent the value function with as few basis functions as possible, thereby reducing the computational complexity of value function approximation and preventing overfitting. The use of basis selection algorithms not only addresses the scaling problem but also allows for tailoring the approximation to a specific policy. This results in a more accurate representation than obtained when using the same subset of basis functions irrespective of the policy being evaluated. To make effective use of the data, we develop a hybrid leastsquares algorithm for setting basis function coefficients. This algorithm is a parametric combination of two common least-squares methods used for MDPs. We provide a geometric and analytical interpretation of these methods and demonstrate the hybrid algorithm’s ability to discover improved policies. We also show how the algorithm can include graphbased regularization to help with sparse samples from stochastic environments. This work investigates all aspects of linear value function approximation: constructing a dictionary of basis functions, selecting a subset of basis functions from the dictionary, and setting the coefficients on the selected basis functions. We empirically evaluate each of these contributions in isolation and in one combined architecture.

Policy gradient is a useful model-free reinforcement learning approach, but it tends to suffer from instability of gradient estimates. In this paper, we analyze and improve the stability of policy gradient methods. We first prove that the variance of gradient estimates in the PGPE (policy gradients with parameter-based exploration) method is smaller than that of the classical REINFORCE method under a mild assumption. We then derive the optimal baseline for PGPE, which contributes to further reducing the variance. We also theoretically show that PGPE with the optimal baseline is more preferable than REINFORCE with the optimal baseline in terms of the variance of gradient estimates. Finally, we demonstrate the usefulness of the improved PGPE method through experiments. 1