Reinforcement learning agents applied to a class of control system problems

Some control system problems are sufficiently complex that it is difficult to define all of the decision weighting up-front. Reinforcement Learning (RL) can be applied to "tune" system performance. The Centipede Game from Economic Game Theory is used to simulate a control system problem with competing subsystem goals. RL agents are applied to the Centipede Game, which pits agents against each other in a game with increasing payoffs if cooperation is developed between the players. The Centipede Game has been studied both in theory and empirically using human players by other researchers. Human players are more inclined to cooperate and achieve the longer-term payoff than economic game theory would predict. This paper focuses on an experimental study of cooperation between RL players (agents) without explicit communications between the players. This work has application in control systems problems where communications is constrained.