Types of Agents — Model-Based Reflex Agents — Beginner-Friendly Notes

1. What Changes vs Simple Reflex

Model-based = reflex plus memory. It keeps an internal model (state) that tracks what the world is like now, even when not directly observed.

Percept + Internal State + Model (how world evolves) → Action

Think: vacuum that remembers "I cleaned A, dirt may still be in B."

2. Inside the Agent (two new parts)

1. State (memory): best guess of current world (e.g., dirt_in_A? dirt_in_B? position?).
2. Model: two pieces of knowledge:
   • Transition model: how actions change the world (move Right from A → at B).
   • Sensor model: how world causes percepts (if dirt_in_current_room → sensor=Dirty).
3. Update: new_state = update(old_state, action, percept)

3. Concrete Example — Vacuum with Memory

• Start: state = {A: unknown, B: unknown}
• Percept A=Dirty → action Suck → model predicts A=Clean → state {A: Clean, B: unknown}
• Next move to B → percept B=Clean → state {A: Clean, B: Clean} → can stop.
  Without memory it would recheck A forever.

4. When You Need It

• Partially observable worlds — need to remember hidden parts.
• Example: self-driving (remember cars occluded behind truck), medical (remember previous tests), game with hidden cards.

5. Still a Reflex at Heart

It still uses condition-action rules, but on internal state not raw percept. No goals or planning yet — just "if my model says X then do Y."

6. One-Line Summary to Write in Exam

A model-based reflex agent maintains an internal state updated by a model of world dynamics and sensor behaviour, allowing it to handle partial observability by reasoning from history rather than the current percept alone.

7. Quick Self-Check

1. What two models does the agent keep?
2. Why is memory enough but planning still missing?