Unity Machine Learning Agents is an open-source toolkit designed to allow anyone develop AI and to integrate machine learning AI into their Unity Engine games. The toolkit can be integrated with and ran through the Unity Development Platform where agents and environments can be configured, allowing developers to design scenarios as if they were creating games. This paper evaluates the possibility of replacing traditional bots with agents and what limitations they have. The agents were trained with the toolkits PPO trainerthrough two scenarios, a puzzle and a maze, where the agent’s information input is limited to artificial vision in 2-dimentions.

The results show that the agents could solve the puzzle and maze, but that they struggle when facing new problems due to the limitations of PPO training when complexity increases unexpectantly. The agents could therefore solve specific problems and they work well as a bot replacement if specialized but cannot at this point replace traditional bots in games where they need specialty in several areas.