Dong-geon Lee and Hyeoncheol Kim, Korea University, Republic of Korea

Applying reinforcement learning agents to the real-world is important. Designing the reward function has problems, especially when it needs to intricately reflect the real-world or requires burden human effort. Under such circumstances, we propose a semi-reward function. This system is intended that each agent can go toward an individual goal when a collective goal is not defined in advance. The semi-reward function, does not require sophisticated reward design, is defined by 'not allowed actions' in the environments without any information about the goal. A tutorial-based agent can sequentially determine actions based on its current state and individual goal. It can be learned through a semi-reward function and toward its own goal. For the combination of these two, we constructed training method to reach the goal. We demonstrate that agents trained in arbitrary environments could go toward it own goal even if they are given different goals in different environments.

Reinforcement Learning, Reward Function, Reward Engineering, Transformer-based Agent, Goal-based Agent