CHALLENGES IN MULTI-AGENT REINFORCEMENT LEARNING: EXPLORATION, COMMUNICATION, AND TRAINING STRATEGY, DATE
Abstract
Despite the recent achievements, MARL faces several challenges that are common across the different application areas. In MARL, agent’s interactions can be regarded as a changing environment, and the joint action spaces generated by such dynamics increase exponentially with the number of agents. In addition, the absence of visibility among agents makes the problem more complex and intractable. This research proposal submitted by ETRI Korea (PI, Sungwon Yi) is intended to explore closely intertwined challenges in MARL. First, we develop acommunication mechanism For MARL as a solution for the visibility problem. The solution will be carefully designed and built upon our previous work within a semantic communication concept in response to the unreliable communication environment. Subsequently, a novel exploration technique will be developed, which is essential in dealing with complex search spaces in MARL. Here, a solution will be developed based on the in-depth analysis of a MARL reward structure, targeting real-world applications. Next, a role-based MARL scheme will be developed. Here, we investigate a technique to train a group of agents to exhibit specific behaviors based on the Grey Wolf Optimizer (GWO) concept to reduce search space and training time. Finally, experimental performance analysis on each proposed technique under a generalization perspective is also planned to be conducted jointly with a research group (PI Hyun Oh Song) from the Seoul National University (SNU).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2023
- Source ID
- FA23862214009
Entities
People
- Sungwon Yi
Organizations
- Air Force Office of Scientific Research
- Electronics and Telecommunications Research Institute
- United States Air Force