2.3 The Properties of Task Environments (BT104CO)
1. Understanding the Task Environment
In the framework of Russell and Norvig, the design of an agent is strictly dictated by the Task Environment. To design a rational agent, we must first categorize the environment using six key dimensions.
Fully vs. Partially Observable
Fully: Sensors detect all relevant aspects (e.g., Chess).
Partially: Missing info due to noise or limited range (e.g., Poker).
Single Agent vs. Multiagent
Single: Acting alone (e.g., Crossword).
Multiagent: Other agents are present. Can be Competitive (Chess) or Cooperative (Traffic).
Deterministic vs. Stochastic
Deterministic: Next state is completely certain (e.g., Sudoku).
Stochastic: Uncertainty or randomness involved (e.g., Taxi driving).
Episodic vs. Sequential
Episodic: Actions are independent "episodes" (e.g., Image Analysis).
Sequential: Decisions have long-term consequences (e.g., Driving).
Static vs. Dynamic
Static: World doesn't change while thinking (e.g., Crossword).
Dynamic: World keeps moving during deliberation (e.g., Taxi).
Discrete vs. Continuous
Discrete: Limited, distinct number of states (e.g., Chess).
Continuous: Smoothly varying speed, angles, coords (e.g., Self-driving cars).
2. Summary Table of Task Environments
| Environment | Observable | Agents | Deterministic | Episodic | Static | Discrete |
|---|---|---|---|---|---|---|
| Chess | Fully | Multi | Strategic | Sequential | Static | Discrete |
| Taxi Driving | Partial | Multi | Stochastic | Sequential | Dynamic | Continuous |
| Medical Diagnosis | Partial | Single | Stochastic | Sequential | Dynamic | Continuous |
| Image Analysis | Fully | Single | Deterministic | Episodic | Semi | Continuous |
3. Why This Matters for AI Design
"Toy Worlds"
Fully Observable, Deterministic, Static, Discrete.
Simple search algorithms and "Simple Reflex" architectures can solve these perfectly.
The Real World
Partially Observable, Stochastic, Dynamic, Continuous.
Requires sophisticated "Model-Based" or "Utility-Based" agents that handle uncertainty.