Sensorless or conformant problem: percepts provide no information at all.
Sensorless plans can save time and cost.
Instead of knowing the state, the agent will have a set of Belief States.
In the belief states, the problem might be fully observable.
Solution is a sequence of actions (not a Conditional Plan)
even if the environment is non-deterministic, as we can’t know what state we end up in.
We can adjust existing problems (P) into belief state problems:
States
every possible subset of underlying states (i.e. $2^n$ states)
Initial state
typically a set of all states in P.
Actions
for b={s_j,...,s_k} ACTIONS(b) = union of all ACTIONS_p(s_i)
this only works if actions are legal from any state, or result in no effect.
Transition Model
RESULTS(b,a) = {s':s' = RESULT_p(s,a) for s in b}
Goal test
IS-GOAL(b) if IS-GOAL_p(s) for every s in b
Action Cost
Complex if actions have different costs from different states, otherwise action cost is the same.
Search Problem