Reliability assessment of multi-agent reinforcement learning algorithms for hybrid local electricity market simulation

The reliability of data-driven multi-agent reinforcement learning (MARL) algorithms is a critical concern, particularly for complex, large-scale multi-agent decision-making problems. This paper aims to assess the reliability of various MARL algorithms in supporting decision-making for prosumer participants in a hybrid local electricity market (LEM) that combines community-based markets and a peer-to-peer (P2P) market. Specifically, it compares the performance of three MARL algorithms: the multi-agent deep deterministic policy gradient (MADDPG) algorithm and two advanced variants incorporating mean-field approximation and attention mechanisms. To evaluate the reliability of these data-driven MARL algorithms, a model-based bi-level optimization problem is introduced for each agent to assess convergence speed and the proximity of results to the ε-Nash equilibrium, as indicated by the no-regret index. The no-regret index is calculated within a mathematical program with equilibrium constraints (MPEC) by fixing the other agents’ behavior generated from the MARL algorithms. Simulation results demonstrate that the attention-MADDPG algorithm achieves the highest no-regret index (0.81), indicating convergence closest to equilibrium, and the greatest total cost reduction (983€), outperforming the other MARL algorithms. The mean-field-MADDPG algorithm is the most balanced, exhibiting robust convergence with the second-highest no-regret index (0.78) and cost reduction (958.8€) under the lowest computational burden (5.4 seconds per episode).