General presentation
The purpose of this activity is to make contact with the autonomous learning agents (ml-agents) of Unity. This activity has been the most complex by far, and I am already saying that the results are not the best, but due to time and hardware limitations I have not been able to give more of myself with this project. Even so, I am proud of what I have learned and I hope that this document reflects it correctly.
Point out that, with respect to ACT3, some of the other agents work a little worse due to the version change I had to make to use ml-agents (from 2022 to 2023).
GitLab repository: https://gitlab.com/jongompal/artificial-intelligence-unity
Repository structure
- ACT2: Unity project of activity 2.
- Artificial Intelligence: Unity project with activity 3 and PAC1 and 2.
- The assets folder includes all the content of the previous activities. For activity 3, files have been added in Scripts, scenes, prefabs and the ml-agents folder has been created with everything related to this topic (including the .onnx files that are worth using)
- The “final” scene is the one called Main V3-ML
- Build: server build attempt for parallel training
- ml-agents: Unity repository to run the training environments, with all the configurations (.yaml) and the results of the multiple (many) attempts. Available in the config and results folders respectively.
Activities
- ACT 1: installation of ML-Agents
- ACT 2: Roller Ball tutorial
- ACT 3: My own ML-Agent
Conclusions
If I had to define this activity with one word, it would be frustration. And it’s not because of the syllabus, even because of my results; This is because I am personally passionate about artificial intelligence (neural network technology more correctly) and I really wanted to do this activity for this exercise. What really frustrated me was burning myself out in the process and not being satisfied with the results.
Definitively, developing models is already a complex task, but doing so in an interactive environment such as that of video games is even more so. My respect to all those who master it and dedicate themselves to it. And I’m sorry for the length of this document, but I felt that I should capture all my experience with this work.