Machine learning techniques are already commonly applied for developing the behaviour of intelligent agents in video games. Most commonly the development of agents is executed via reinforced learning, a relatively simple approach, capable of producing robust solutions to various learning challenges. In the presented thesis we tested whether genetic algorithms could be a viable alternative to reinforced learning. Even though genetic algorithms are very simple and easy to implement they have not seen much use when it comes to development of intelligent agents. To compare the quality of our genetic algorithms based solution, we compared it with ML-Agents, a widespread framework for development of intelligent agents, based on reinforced learning. The comparison of both learning methods. was executed on two popular games under comparable conditions. Our results suggest that genetic algorithms could represent a viable alternative to reinforced learning, but only in simple scenarios. When applied to more complex scenarios, our implementation of genetic results did not fare extremely well.
|
