Enhancing Gameplay Experience Through Reinforcement Learning in Games

It is important to improve the game development field since games have great influence on economics and society. One of the ways to enhance game experience is to replace non-playable characters (NPC) by machine learning (ML) agents. Artificial Intelligence (AI) trained by Reinforcement learning algorithms can adapt to the game environment and act based on users’ actions or changes in surroundings. This research project aims to test the idea of replacing the NPC by artificial intelligence agent trained by Reinforcement learning algorithms. The data for this project was collected by observation of experiments inside of the game created in Unity platform. Agent was trained in special toolkit for Unity platform called "Unity ML-Agents". The behavior of the agents was subjected to preparatory measures for replication, including preliminary data processing using behavior cloning methods. The selected algorithm of the Reinforcement learning was the Proximal Policy Optimization, which is based on balance between new actions and its rewards. The training progress of the agents was monitored utilizing TensorBoard, the default package within the Unity ML Agents toolkit. During the research initialization phase, there were set null hypothesis, which states that pre-scripted bots are better for games and alternative hypothesis, which states that AI can replace NPC and enhance the game experience for a player.