Parametrized Deep Q-Networks Learning: Reinforcement Learning with Discrete-Continuous Hybrid Action Space
This addresses a practical problem in reinforcement learning for domains like gaming and robotics where actions are hybrid, though it appears incremental as it builds on existing DRL methods.
The paper tackles the problem of reinforcement learning with discrete-continuous hybrid action spaces, which are common in applications like computer games, by proposing a parametrized deep Q-network (P-DQN) framework that integrates DQN and DDPG without approximation or relaxation. Empirical results on simulated RoboCup soccer and the game King of Glory validate its efficiency and effectiveness.
Most existing deep reinforcement learning (DRL) frameworks consider either discrete action space or continuous action space solely. Motivated by applications in computer games, we consider the scenario with discrete-continuous hybrid action space. To handle hybrid action space, previous works either approximate the hybrid space by discretization, or relax it into a continuous set. In this paper, we propose a parametrized deep Q-network (P- DQN) framework for the hybrid action space without approximation or relaxation. Our algorithm combines the spirits of both DQN (dealing with discrete action space) and DDPG (dealing with continuous action space) by seamlessly integrating them. Empirical results on a simulation example, scoring a goal in simulated RoboCup soccer and the solo mode in game King of Glory (KOG) validate the efficiency and effectiveness of our method.