Goal Recognition over POMDPs: Inferring the Intention of a POMDP Agent
Miquel Ramirez and Hector Geffner
Plan recognition is the problem of inferring the goals and plans of an agent from partial observations of his behavior. Recently, it has been shown that the problem can be formulated and solved using planners, thus reducing plan recognition to plan generation. In this work, we extend the model-based approach to plan recognition to POMDP settings, where actions are stochastic and states are partially observable. The task is to infer a probability distribution over the possible goals of an agent whose behavior results from a POMDP model. The model is shared between agent and observer except for the true goal of the agent that is hidden to the observer. The observations are action sequences that may contain gaps as some or even most of the actions done by the agent are not observed. We show that the posterior goal distribution can be computed from the value function $V_G(b)$ over beliefs $b$ generated by a POMDP planner for each possible goal $G$. More precisely, executions are sampled from this value function assuming that the agent tends to the select the actions that look best, and the probability of the observations $O$ given the goal $G$ is inferred from these samples. This basic scheme is then extended to situations where the agent prior beliefs are not fully known to the observer, or when the observations of the actions done by the agent are corrupted by noise. Extensive experimental results are also reported.