[3J4-GS-5] Agents:

[3J4-GS-5-03] Using Curiosity-driven Search to adapt to changes in prey behavior for Multi-agent Pursuit of Learning Preys

〇Ryosuke Hayashi¹, Koichi Moriyama¹, Atsuko Mutoh¹, Kosuke Shima¹, Tohgoroh Matsui², Nobuhiro Inuzuka¹ (1. Nagoya Institute of Technology, 2. Chubu University)

Keywords:Multi-Agent, Reinforcement learning

Multi-agent pursuit problems are difficult to learn with reinforcement learning due to few chance of getting rewards. Curiosity-driven search that gives intrinsic rewards based on familiarity of states is a promising method but those depending on a fixed network like RND do not detect a slight but important change, e.g., behavior of preys, in the state expressed as a high-dimensional vector. This work propose to use SND, a novel curiosity-driven search method which first learns the target network to separate its outputs depending on such important elements in the state vectors. Once we learn the target network to separate its outputs based on the velocities of preys, the predators are given intrinsic rewards from a slight change of them and encouraged to explore such states. We compare the result of the proposal with that of an existing work using RND.