Reinforcement Learning Agent Design by Using a Finite State Machine with Deep Neural Networks

Abstract

In this study, a model that uses a finite state machine with deep neural network to control reinforcement learning is supposed. One trained deep neural network is set up for each state, and these neural networks switch depending on the state transition to achieve the control of the character's movement. We built a state machine with four states in the Unity3D environment, and implemented the movement of the character that takes the ball in the right side area and transports it to the goal in the left side area by using reinforcement learning.