Foundational Aspects of Machine Learning in Multi-Agent Online Games as Serious Games
Abstract
This seedling project investigated the complexities of scaling reinforcement learning algorithms, using commercial-off-the-shelf strategy games as the experimental environment. A major research contribution is an multi-agent reinforcement learning (MARL) approach where an agent learns policies based on other agents, rather than attempting to learn policy by itself, independent of the other agents. The research team also investigated using the Grey Wolf Optimizer simulate human-AI teaming and training, as a simpler method to approximate training conditions with hybrid teams. The project produced three papers and two patent submissions. In addition, software developed for the experiment is openly hosted on GitHub.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 05, 2021
- Accession Number
- AD1152078
Entities
People
- Sungwon Yi
Organizations
- Electronics and Telecommunications Research Institute