Reinforcement Learning ve hře žánru 2D platformer v prostředí Unity

Abstract

This bachelor’s thesis presents the design, implementation and evaluation of a system for training an autonomous agent in a two-dimensional platformer game environment using deep reinforcement learning algorithms. Unity game engine combined with the ML-Agents framework was chosen as the simulation and development platform, providing communication between the game environment and a PyTorch-based optimization backend. The thesis compares two modern algorithms Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) on an identical environment with progressively increasing difficulty. Potential-Based Reward Shaping was implemented to address the sparse reward problem, and Curriculum Learning with weight transfer between training phases was used to facilitate efficient training across five levels. Experimental results show that PPO achieved high success rates on four out of five levels, while SAC succeeded only on the first two. The thesis demonstrates that algorithm selection has a significant impact on curriculum learning success and that the on-policy nature of PPO makes it more suitable for this specific environment.