Teaching a machine learning agent to survive in a 2D top-down environment
Ropilo, Teemu (2019)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2019120925566
https://urn.fi/URN:NBN:fi:amk-2019120925566
Tiivistelmä
Machine learning is not a new thing, but the rapid development of computer architecture has brought the tools in the hands of regular users. At the same time, this has led to advances in machine learning assisted game development, introducing new elements to traditional scripted gameplay behaviour.
The objective of this thesis was to train an autonomous machine learning agent that can survive inside a two-dimensional top-down environment, while performing the tasks required to reach the goals set for it. In addition to training the agent, different learning methods were examined and compared, to find the best method for training the agent.
Before the training environment was created, references were gathered, related to machine learning, especially in the area of game development. The preliminary design for the implementation was then created, which was followed by outlining the documentation. To start developing the training environment, the machine learning tools were configured, and initial tests were made, as a proof of concept. At this point all the necessary elements were present, so from this point forward the process continued with improving on each element. During the training of the machine learning agent, statistics were gathered to examine the efficiency of each learning method. Finally, the statistics were compared alongside the machine learning models, to find on the best learning method for the purposes of this thesis.
The findings showed that imitation learning assisted reinforcement learning was the best learning methods. Even though further tuning of the parameters and environment could have given better results, the obtained results were deemed a success. Reinforcement learning and curriculum learning methods also showed promise, but they were not as efficient. The most significant result of this thesis was the accumulated knowledge on the subject, opening new possibilities in the area of machine learning and game development.
The objective of this thesis was to train an autonomous machine learning agent that can survive inside a two-dimensional top-down environment, while performing the tasks required to reach the goals set for it. In addition to training the agent, different learning methods were examined and compared, to find the best method for training the agent.
Before the training environment was created, references were gathered, related to machine learning, especially in the area of game development. The preliminary design for the implementation was then created, which was followed by outlining the documentation. To start developing the training environment, the machine learning tools were configured, and initial tests were made, as a proof of concept. At this point all the necessary elements were present, so from this point forward the process continued with improving on each element. During the training of the machine learning agent, statistics were gathered to examine the efficiency of each learning method. Finally, the statistics were compared alongside the machine learning models, to find on the best learning method for the purposes of this thesis.
The findings showed that imitation learning assisted reinforcement learning was the best learning methods. Even though further tuning of the parameters and environment could have given better results, the obtained results were deemed a success. Reinforcement learning and curriculum learning methods also showed promise, but they were not as efficient. The most significant result of this thesis was the accumulated knowledge on the subject, opening new possibilities in the area of machine learning and game development.