5DV248 - Lab 2: Rationality depends on the performance measure

5DV248 Lecture 1 · Move the sliders (the performance measure), press Evaluate, and watch the ranking of the agents change.

Performance measure (the designer's choice) Reward per clean square at the end: 10 Penalty per move: 1 Penalty per Suck: 0.5 Penalty per dirty cell per time step: 0 Probability dirt respawns: 0.05

Agents (all simple reflex, differing in rules):
reflex: Suck if dirty, else cross to the other square (slides).
paranoid: same, but keeps moving even when all is clean.
lazy: Suck if dirty, else NoOp (never moves).
random: uniform random action.

Average performance (100 runs, horizon 20 steps)

Lab 2: The same agent, different performance measures