POMDPs for Facial Display Understanding

It is a tuple S,A,T,R,O,B

• S: finite set of unobservable states
• A: finite set of agent actions
• T: S x A -> S transition function
• R: S x A -> R reward function
• O: set of observations
• B: S x A -> O observation function

A POMDP is usually used for an agent in a simple reactive environment. If her environment includes other agents, the situation becomes a bit more complicated. However, if we take a decision-analytic approach to decisions, then each agent can use a POMDP to choose strategies if she includes the strategies, value functions, and internal states of all other agents as part of her internal state.

Since we wish to focus on non-verbal displays, we consider that only these are non-observable. The model is then as shown here

The observation function, P(O|A), gives the probability of making some observation (a complete video sequence), given a high-level description of the facial display or gesture. This is a computer vision problem.

Read on about this observation function,