Types of Environments — Observable vs Partially Observable — Beginner-Friendly Notes
1. Core Idea
Observability = can the agent see everything it needs to decide?
- Fully observable: sensors give the complete, accurate state at each moment. Agent does not need memory to know what's happening.
- Partially observable: sensors are limited, noisy, or hidden — agent must guess or remember.
This is the most important environment property — it decides if you need memory/models.
2. Everyday Examples
| Environment | Observability | Why? |
|---|---|---|
| Chess | Fully observable | Whole board visible |
| Vacuum robot with camera seeing whole room | Fully observable | All dirt and obstacles seen |
| Vacuum with only local dirt sensor | Partially observable | Does not know dirt in other room |
| Self-driving car | Partially observable | Occluded cars, noisy sensors, can't see driver intent |
| Poker | Partially observable | Opponent cards hidden |
| Medical diagnosis | Partially observable | Hidden disease state, only symptoms visible |
3. What Agent Needs in Each Case
- Fully observable: can be simple — decide from current percept alone (reflex agents can work).
- Partially observable: needs internal model / memory — keep a belief state ("maybe dirt in B"), update it with each percept. This is why we need model-based agents (coming topic).
4. Partial Observability Comes From Three Causes
- Hidden state — relevant info not sensed (cards, other room).
- Noisy sensors — inaccurate readings (camera blur, GPS error).
- Limited sensors — agent only senses nearby.
5. Quick Test in Exam
Ask: "Can the agent see all of what affects its decision right now without remembering past?"
- Yes → fully observable
- No → partially observable
6. One-Line Summary to Write in Exam
An environment is fully observable when the agent's sensors reveal the complete state at each step; otherwise it is partially observable and the agent must maintain an internal model and use memory.
7. Quick Self-Check
- Is a maze where you see only your cell fully observable?
- Why is chess fully observable but poker not?
- What extra component does a partially observable agent need?