Reinforcement Learning Repository at MSU

Topics: Partially Observable Problems

The vast majority of RL work concentrates on the completely observable case, in which there is noise in the agent's actions, but none in its observation of the environment. When the state space is not completely observable, additional methods need to be used beyond those which apply to MDPs. Partially observable Markov decision processes (POMDP's) extend the MDP model to include an observational model (which is a stochastic function of the current state). The true state is hidden because the number of observations is typically far less than the number of possible states.