Parametrized Deep Q-networks Learning: Playing Online Battle Arena with Discrete- Continuous Hybrid Action Space

Most existing deep reinforcement learning (DRL) frameworks consider action spaces that are either discrete or continuous space. Motivated by the project of design Game AI for King of Glory (KOG), one the world’s most popular mobile game, we consider the scenario with the discrete-continuous hybrid action space. To directly apply existing DLR frameworks, existing approaches either approximate the hybrid space by a discrete set or relaxing it into a continuous set, which is usually less efficient and robust. In this paper, we propose a parametrized deep Q-network (P-DQN) farmework for the hybrid action space without approximation or relaxation. Our algorithm combines DQN and DDPG and can be viewed as an extension of the DQN to hybrid actions. The empirical study on the game KOG validates the efficiency and effectiveness of our method.