Foundational Aspects of Machine Learning in Multi Agent Online Games as Serious Games
Abstract
This one year seedling proposal by ETRI Korea is intended to build an experimental framework for a complex multi agent role playing game which enables an AI to participate, observe and act. The AI’s actions will be based on a dynamically trained Deep Neural Network. The experimental framework to be built will be based on open source Machine Learning frameworks and will support multiple concurrent executions of the selected game to collect game data and provide interfaces for dynamically trained DNN to actuate actions needed in the game by leveraging the open sourced APIs. ETRI will experiment with Defense of the Ancients II, DOTA2. In the one year seedling, we will carefully explore the various design choices made in the creation of OpenAI5 for DOTA2. A novel element of our research is to reduce the computational burden of training AIs in complex games, for example by storing prior experience. Such a repository can help reduce the amount of future exploration of play action-strategy spaces by purely random moves. Also, a novel element of our proposed research will be to explore how best to form hybrid AI-human players. That is, we will compare the advantages of fully AI and fully human players to hypothesize how best to form hybrid players. ETRI will provide cost share for this project so that the seedling grant will support at least two engineers. This seedling effort will be assisted by the Korean company RTST who has been developed machine learning based training programs, games. Furthermore, this seedling effort will be supported by a team from the Pennsylvania State University which is funded in part by an existing AFOSR DDDAS grant. The PSU ETRI team will produce literature surveys during the seedling year on AIs of competitive games, and also that of interpretable-explainable machine learning, long term.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 14, 2022
- Source ID
- FA23861914020XX0
Entities
People
- Sungwon Yi
Organizations
- Air Force Office of Scientific Research
- Electronics and Telecommunications Research Institute
- United States Air Force