Problems where there is limited sensing of the environment:
- a sliding tile puzzle with just one corner visible
- a local sensing vacuum
For every action:
RESULTS(b, a) = {b_0: b_0 = UPDATE(PREDICT(b,a), o) and o belongs to POSSIBLE-PERCEPTS(PREDICT(b,a))}
Transition in local-sensing vacuum world:
- Action results in two belief states
- Sensing narrows the possible states
[Suck, Right, if Rstate = {6} then Suck else []]
Agents in partially observable environments:
- Solution will be a conditional plan (not a path or sequence)
- The agent will maintain a Belief State
b' = UPDATE(PREDICT(b,a),o)- Monitoring, filtering or state estimation
Rationality