State Aggregation for Distributed Value Iteration in Dynamic Programming

We propose a distributed algorithm to solve dynamic programming problem with multiple agents, where each agent has only partial knowledge of the state transition probabilities and costs. provide consensus proofs for presented derive error bounds obtained value function respect what is considered as "true solution" from conventional iteration. To minimize communication overhead between costs are aggregated shared agents when updated expected influence solution other significantly. demonstrate efficacy proposed aggregation method large-scale urban traffic routing problem. Individual compute fastest route common access point share local congestion information allowing fully minimal agents.