Improved Regret Bound and Experience Replay in Regularized Policy Iteration
arXiv:2102.12611
Abstract
In this work, we study algorithms for learning in infinite-horizon undiscounted Markov decision processes (MDPs) with function approximation. We first show that the regret analysis of the Politex algorithm (a version of regularized policy iteration) can be sharpened from to under nearly identical assumptions, and instantiate the bound with linear function approximation. Our result provides the first high-probability regret bound for a computationally efficient algorithm in this setting. The exact implementation of Politex with neural network function approximation is inefficient in terms of memory and computation. Since our analysis suggests that we need to approximate the average of the action-value functions of past policies well, we propose a simple efficient implementation where we train a single Q-function on a replay buffer with past data. We show that this often leads to superior performance over other implementation choices, especially in terms of wall-clock time. Our work also provides a novel theoretical justification for using experience replay within policy iteration algorithms.
References in corpus (12)
- Overcoming catastrophic forgetting in neural networks
- Playing Atari with Deep Reinforcement Learning
- Trust Region Policy Optimization
- Continual Learning Through Synaptic Intelligence
- DeepMind Control Suite
- REGAL: A Regularization based Algorithm for Reinforcement Learning in Weakly Communicating MDPs
- Coresets via Bilevel Optimization for Continual Learning and Streaming
- A Short Note on Concentration Inequalities for Random Vectors with SubGaussian Norm
- Mirror Descent Policy Optimization
- Adaptive Approximate Policy Iteration
- Learning Infinite-horizon Average-reward MDPs with Linear Function Approximation
- Quinoa: a Q-function You Infer Normalized Over Actions