Collaborating Authors

Bhatnagar, Shalabh

Multi-Step Dyna Planning for Policy Evaluation and Control

Neural Information Processing Systems

We extend Dyna planning architecture for policy evaluation and control in two significant aspects. First, we introduce a multi-step Dyna planning that projects the simulated state/feature many steps into the future. Our multi-step Dyna is based on a multi-step model, which we call the {\em $\lambda$-model}. The $\lambda$-model interpolates between the one-step model and an infinite-step model, and can be learned efficiently online. Second, we use for Dyna control a dynamic multi-step model that is able to predict the results of a sequence of greedy actions and track the optimal policy in the long run.

Convergent Temporal-Difference Learning with Arbitrary Smooth Function Approximation

Neural Information Processing Systems

We introduce the first temporal-difference learning algorithms that converge with smooth value function approximators, such as neural networks. Conventional temporal-difference (TD) methods, such as TD($\lambda$), Q-learning and Sarsa have been used successfully with function approximation in many applications. However, it is well known that off-policy sampling, as well as nonlinear function approximation, can cause these algorithms to become unstable (i.e., the parameters of the approximator may diverge). Sutton et al (2009a,b) solved the problem of off-policy learning with linear TD algorithms by introducing a new objective function, related to the Bellman-error, and algorithms that perform stochastic gradient-descent on this function. In this paper, we generalize their work to nonlinear function approximation.

Universal Option Models

Neural Information Processing Systems

We consider the problem of learning models of options for real-time abstract planning, in the setting where reward functions can be specified at any time and their expected returns must be efficiently computed. We introduce a new model for an option that is independent of any reward function, called the {\it universal option model (UOM)}. We prove that the UOM of an option can construct a traditional option model given a reward function, and the option-conditional return is computed directly by a single dot-product of the UOM with the reward function. We extend the UOM to linear function approximation, and we show it gives the TD solution of option returns and value functions of policies over options. We provide a stochastic approximation algorithm for incrementally learning UOMs from data and prove its consistency.

Hierarchical Average Reward Policy Gradient Algorithms Artificial Intelligence

Option-critic learning is a general-purpose reinforcement learning (RL) framework that aims to address the issue of long term credit assignment by leveraging temporal abstractions. However, when dealing with extended timescales, discounting future rewards can lead to incorrect credit assignments. In this work, we address this issue by extending the hierarchical option-critic policy gradient theorem for the average reward criterion. Our proposed framework aims to maximize the long-term reward obtained in the steady-state of the Markov chain defined by the agent's policy. Furthermore, we use an ordinary differential equation based approach for our convergence analysis and prove that the parameters of the intra-option policies, termination functions, and value functions, converge to their corresponding optimal values, with probability one. Finally, we illustrate the competitive advantage of learning options, in the average reward setting, on a grid-world environment with sparse rewards.

A Convergent Off-Policy Temporal Difference Algorithm Machine Learning

Learning the value function of a given policy (target policy) from the data samples obtained from a different policy (behavior policy) is an important problem in Reinforcement Learning (RL). This problem is studied under the setting of off-policy prediction. Temporal Difference (TD) learning algorithms are a popular class of algorithms for solving the prediction problem. TD algorithms with linear function approximation are shown to be convergent when the samples are generated from the target policy (known as on-policy prediction). However, it has been well established in the literature that off-policy TD algorithms under linear function approximation diverge. In this work, we propose a convergent on-line off-policy TD algorithm under linear function approximation. The main idea is to penalize the updates of the algorithm in a way as to ensure convergence of the iterates. We provide a convergence analysis of our algorithm. Through numerical evaluations, we further demonstrate the effectiveness of our algorithm.

Generalized Speedy Q-learning Artificial Intelligence

In this paper, we derive a generalization of the Speedy Q-learning (SQL) algorithm that was proposed in the Reinforcement Learning (RL) literature to handle slow convergence of Watkins' Q-learning. In most RL algorithms such as Q-learning, the Bellman equation and the Bellman operator play an important role. It is possible to generalize the Bellman operator using the technique of successive relaxation. We use the generalized Bellman operator to derive a simple and efficient family of algorithms called Generalized Speedy Q-learning (GSQL-w) and analyze its finite time performance. We show that GSQL-w has an improved finite time performance bound compared to SQL for the case when the relaxation parameter w is greater than 1. This improvement is a consequence of the contraction factor of the generalized Bellman operator being less than that of the standard Bellman operator. Numerical experiments are provided to demonstrate the empirical performance of the GSQL-w algorithm.

Solution of Two-Player Zero-Sum Game by Successive Relaxation Machine Learning

We consider the problem of two-player zero-sum game. In this setting, there are two agents working against each other. Both the agents observe the same state and the objective of the agents is to compute a strategy profile that maximizes their rewards. However, the reward of the second agent is negative of reward obtained by the first agent. Therefore, the objective of the second agent is to minimize the total reward obtained by the first agent. This problem is formulated as a min-max Markov game in the literature. The solution of this game, which is the max-min reward (of first player), starting from a given state is called the equilibrium value of the state. In this work, we compute the solution of the two-player zero-sum game utilizing the technique of successive relaxation. Successive relaxation has been successfully applied in the literature to compute a faster value iteration algorithm in the context of Markov Decision Processes. We extend the concept of successive relaxation to the two-player zero-sum games. We prove that, under a special structure, this technique computes the optimal solution faster than the techniques in the literature. We then derive a generalized minimax Q-learning algorithm that computes the optimal policy when the model information is not known. Finally, we prove the convergence of the proposed generalized minimax Q-learning algorithm.

Second Order Value Iteration in Reinforcement Learning Machine Learning

Value iteration is a fixed point iteration technique utilized to obtain the optimal value function and policy in a discounted reward Markov Decision Process (MDP). Here, a contraction operator is constructed and applied repeatedly to arrive at the optimal solution. Value iteration is a first order method and therefore it may take a large number of iterations to converge to the optimal solution. In this work, we propose a novel second order value iteration procedure based on the Newton-Raphson method. We first construct a modified contraction operator and then apply Newton-Raphson method to arrive at our algorithm. We prove the global convergence of our algorithm to the optimal solution and show the second order convergence. Through experiments, we demonstrate the effectiveness of our proposed approach.

Reinforcement Learning in Non-Stationary Environments Machine Learning

Reinforcement learning (RL) methods learn optimal decisions in the presence of a stationary environment. However, the stationary assumption on the environment is very restrictive. In many real world problems like traffic signal control, robotic applications, one often encounters situations with non-stationary environments and in these scenarios, RL methods yield sub-optimal decisions. In this paper, we thus consider the problem of developing RL methods that obtain optimal decisions in a non-stationary environment. The goal of this problem is to maximize the long-term discounted reward achieved when the underlying model of the environment changes over time. To achieve this, we first adapt a change point algorithm to detect change in the statistics of the environment and then develop an RL algorithm that maximizes the long-run reward accrued. We illustrate that our change point method detects change in the model of the environment effectively and thus facilitates the RL algorithm in maximizing the long-run reward. We further validate the effectiveness of the proposed solution on non-stationary random Markov decision processes, a sensor energy management problem and a traffic signal control problem.

Successive Over Relaxation Q-Learning Machine Learning

In a discounted reward Markov Decision Process (MDP) the objective is to find the optimal value function, i.e., the value function corresponding to an optimal policy. This problem reduces to solving a functional equation known as the Bellman equation and a fixed point iteration scheme known as the value iteration is utilized to obtain the solution. In [1], a successive over-relaxation based value iteration scheme is proposed to speed up the computation of the optimal value function. They propose a modified Bellman equation and prove faster convergence to the optimal value function. However, in many practical applications, the model information is not known and we resort to Reinforcement Learning (RL) algorithms to obtain optimal policy and value function. One such popular algorithm is Q-Learning. In this paper, we propose Successive Over Relaxation (SOR) Q-Learning. We first derive a fixed point iteration for optimal Q-values based on [1] and utilize stochastic approximation to derive a learning algorithm to compute the optimal value function and an optimal policy. We then prove the convergence of the SOR Q-Learning to optimal Q-values. Finally, through numerical experiments, we show that SOR Q-Learning is faster compared to the standard Q-Learning algorithm.