Study on the convergence property of re learning model in electricity market simulation

Electricity market being too complex to be modeled by standard microeconomic and game theoretic approaches, agent-based computational economics (ACE) plays more and more important role in electricity market study. In this paper, the convergence property of Roth-Erev (RE) reinforcement learning method in electricity market simulation is studied. Simulation results based on a 4-generator system are presented. The results demonstrate that the convergence period and convergence price are effected by many factors, such as the pseudorandom number generator and the parameter k. Overall, the clearing price converge to is inversely proportional to the period number converge at, which indicates the contradiction in the calibration of parameter k . The reason of the contradiction is analyzed from the mechanism of reinforcement learning.