Types of Agents — Utility-Based Agents — Beginner-Friendly Notes

1. What Changes vs Goal-Based

Utility-based = goal-based plus a measure of how good a state is. Not just "reach goal or not", but "which path is better, safer, cheaper, faster?"

State + Utility Function → Choose action that maximises Expected Utility

Think: GPS with preferences — fastest vs shortest vs most fuel-efficient.

2. Utility — Simple Idea

• Utility $U(s)$: a number scoring how desirable a state is (higher = better). Can encode safety, cost, time, happiness.
• Expected utility: when outcomes are uncertain, agent calculates weighted average: $EU(a) = \Sigma P(result|a) \cdot U(result)$.
• Example: two paths to hospital:
  • Path A: 10 min, risky road → $U=5$
  • Path B: 15 min, safe road → $U=9$ → utility-based picks B even though both reach goal.

3. Why Goals Are Not Enough

• Multiple successful plans: which to pick? Goal agent picks any; utility agent picks best.
• Conflicting goals: safety vs speed — utility trades them off numerically.
• Uncertainty: when action may succeed or fail, utility chooses best expected outcome (links to Bayes, Module 4).

4. Economics Link (for viva)

Utility comes from economics — rational agent = maximises utility, not just achieves a goal. That's why AI adopts it as the definition of rationality.

5. Quick Comparison

Agent Type	Decision Rule
Simple reflex	if percept → action
Model-based	if model-state → action
Goal-based	if action leads to goal → do it
Utility-based	if action leads to highest utility → do it
Learning	improves any of the above from experience

6. Example — Self-Driving Taxi

• Utility combines: U = -time - 10 \cdot risk - fuel_cost + comfort
• Agent evaluates each manoeuvre's expected utility and picks max.

7. One-Line Summary to Write in Exam

A utility-based agent selects actions by maximising a numeric utility (or expected utility under uncertainty), allowing trade-offs between competing objectives and choosing the best among multiple goal-achieving plans.

8. Quick Self-Check

1. Give a state where two paths both reach the goal but have different utilities.
2. Why does utility help under uncertainty?