Comparison of Two Learning Algorithms in Modelling the Generator's Learning Abilities

摘要

This paper discusses the generator's optimal bidding problem. Reinforcement learning is employed to model the generator's learning ability. Through the repeated learning, the generator can develop optimal bidding in the point view of long term. Simulation result shows the generator equipped with learning ability can definitely perform better than the one without learning ability. The learning ability increases the profit of this 'smart' generator, who exercises more market power than the 'normal' generator. It's the main advantage that the generator with learning gets. We compare two learning algorithms, and conclude that SA-Q agent can always converge to the optimal action, but VRE can't. VRE has serious stochastic characteristic, which lead agent converge to one action randomly. In this work, VRE is quite sensitive to the parameters of system. However SA-Q has no such problem, which can lead agent converge to the optimal action.