Exhibit 03Decision Under Uncertainty
Wumpus World
A dark 4×4 cave hides pits, a Wumpus and a pile of gold. Code turns breezes and stenches into risk estimates. Jev sees its own cell and the risks of its neighbours; the classic agent also plans routes across the whole map it has explored, so the two do not see the same amount.
Partly in scopeCode does the inference, but Jev must weigh numeric risk estimates, a documented weakness. TypeSafe docs, Jev 1.13 jaggedness: “It struggles with tasks that require numeric precision.”
World
→State
→Decision
→Action
→New state
World
Step
0 / 40
Score
0
Arrow
1
Gold
not found
+1000 for leaving with the gold, −1000 for dying, −1 per action, −10 for the arrow. Every world here is guaranteed solvable: a safe path to the gold exists.
This world · AIMA Fig. 7.2
Classicnot run yet
Jevnot run yet
Legend
pit risk Wumpus risk
proven safe path taken
Breeze: a pit is next door. Stench: the Wumpus is next door. Glitter: gold is here.
Current percepts
Breeze
NO
Stench
NO
Glitter
NO
Possible next cells
northknown safe
eastknown safe
Jev decision
idlePress Step to send the agent's knowledge to Jev.
MOVE NORTH—
MOVE EAST—
SHOOT NORTH—
SHOOT EAST—
RETREAT—
Lesson · Uncertainty
Beliefs → Decision → New evidence
The work is split on purpose. Ordinary code keeps the hidden world, turns each breeze and stench into pit and Wumpus probabilities by enumerating every consistent possibility (exact under the model's assumptions: independent pits, a uniform prior over where the Wumpus is), and refuses illegal moves. Jev never sees the map, only the agent's own percepts, visit counts and the risk of each neighbouring cell, and picks one action with no memory beyond that snapshot. The classic agent reads the same beliefs but also plans multi-step routes over every cell it has explored, so it has more context than Jev. Differences in outcome mix judgement with that difference in view.