On June 29, 2023, Peter Atrazhev, TREX-Ai’s CFO, reached a significant milestone by successfully defending his master’s thesis titled “Multi-Agent Deep Reinforcement Learning for Autonomous Energy Coordination in Demand Response Methods for Residential Distribution Networks.” This research delves deep into the realm of centralized learning decentralized execution (CLDE) for training deep reinforcement learning agents.
Through rigorous empirical analysis in the Level Based Foraging (LBF) environment and the CityLearn challenge 2022, Peter’s study offered pivotal insights:
-
- Different CLDE algorithms, such as MAPPO and QMIX, vary unpredictably when transitioning from joint to individual rewards in the LBF environment.
-
- MAA2C with individual rewards stood out as highly effective in the CityLearn challenge, excelling in peak demand and district ramping KPIs.
These findings emphasize the nuanced decisions required when selecting between individual and joint rewards in Multi-Agent Reinforcement Learning (MARL) environments. Peter’s insights challenge the default leanings towards joint rewards, suggesting that individual rewards can sometimes outperform in collaborative tasks. Peter’s thesis is available here.
At TREX-Ai, Peter’s research directly influences our technological advancements, particularly in coordination mechanisms. As we continue our journey in artificial intelligence, we’re proud to have such groundbreaking research underpinning our efforts.
Congratulations to Peter on this stellar achievement, which promises to shape the AI landscape for years to come!