Reinforcement Learning Method with Internal World Model Training

Reinforcement learning can be applied to various tasks such as developing motion controllers for humanoid or insect-like robots, playing board games and video games. However, reinforcement learning requires a large number of trials for learning appropriate behaviors in an environment that required physics simulations due to the computational costs of simulations. We propose a model inspired by human brain simulation. The agent builds the internal model based on the environment and trains in a simulated environment by the internal model in addition to training in the actual environment. As a result, the agent acquires the walking movement with fewer trials.