Accelerated Target Updates for Q-learning
arXiv:1905.02841
Abstract
This paper studies accelerations in Q-learning algorithms. We propose an accelerated target update scheme by incorporating the historical iterates of Q functions. The idea is conceptually inspired by the momentum-based accelerated methods in the optimization theory. Conditions under which the proposed accelerated algorithms converge are established. The algorithms are validated using commonly adopted testing problems in reinforcement learning, including the FrozenLake grid world game, two discrete-time LQR problems from the Deepmind Control Suite, and the Atari 2600 games. Simulation results show that the proposed accelerated algorithms can improve the convergence performance compared with the vanilla Q-learning algorithm.
We need further adjustment of some parts of the papaer
References in corpus (9)
- Adam: A Method for Stochastic Optimization
- Playing Atari with Deep Reinforcement Learning
- Prioritized Experience Replay
- DeepMind Control Suite
- Action-Conditional Video Prediction using Deep Networks in Atari Games
- Convergence Analysis of Two-layer Neural Networks with ReLU Activation
- Learning and Generalization in Overparameterized Neural Networks, Going Beyond Two Layers
- Fastest Convergence for Q-learning
- Reinforcement Learning Meets Hybrid Zero Dynamics: A Case Study for RABBIT