Planning, uncertainty and expert systems notes — Unit 5

Planning is the task of coming up with a sequence of actions that will achieve a goal. It bridges the gap between Search (finding a path) and Logic (reasoning about the world).

⇒1.1 STRIPS Representation

The Stanford Research Institute Problem Solver (STRIPS) is the classic language for expressing planning problems. An action is defined by three things:

• Action Name and Parameters: e.g., `Fly(plane, from, to)`.
• Preconditions: What must be true before the action. e.g., `At(plane, from)`.
• Effects: What changes after the action. e.g., `¬At(plane, from) ∧ At(plane, to)`.

In the real world, agents almost never have perfect information. Logic breaks down when we don't know facts with 100% certainty.

⇒2.1 Why Logic Fails

We cannot write the rule `Toothache ⇒ Cavity` because it's not always true. We could write `Toothache ⇒ Cavity ∨ GumDisease ∨ ImpactedWisdomTooth`, but listing every single possibility is impossible.

⇒2.2 The Solution: Probability

Instead of True/False, we assign a degree of belief from 0 to 1. E.g., `P(Cavity | Toothache) = 0.8`. We use Probability Theory to reason under uncertainty.

The foundation of modern probabilistic AI. It allows us to calculate the probability of a cause given an observed effect.

⇒3.1 The Equation

`P(Cause | Effect) = [ P(Effect | Cause) * P(Cause) ] / P(Effect)`

If a patient comes in with a stiff neck (Effect), what is the probability they have Meningitis (Cause)? We calculate this using the known statistics of Meningitis.

A Bayesian Network is a Directed Acyclic Graph (DAG) used to represent the conditional dependencies between different variables.

• Nodes: Represent random variables (e.g., `Burglary`, `Earthquake`, `Alarm`, `JohnCalls`).
• Edges: Represent direct influence (Cause -> Effect).

By utilizing conditional independence (the fact that John calling the police is independent of the Earthquake, given that the alarm went off), Bayesian Networks drastically reduce the amount of math required to compute complex probabilities.

An Expert System is a computer program that emulates the decision-making ability of a human expert.

⇒5.1 Architecture

• Knowledge Base: A massive database of 'IF-THEN' rules extracted from human experts (e.g., Doctors) by Knowledge Engineers.
• Inference Engine: Usually a Backward-Chaining engine that asks the user questions to prove a hypothesis.
• User Interface: How the non-expert user queries the system.

⇒5.2 Famous Example: MYCIN

Developed in the 1970s at Stanford, MYCIN diagnosed bacterial infections. It asked the doctor questions and recommended antibiotics. It actually outperformed junior doctors in accuracy, though it was never used in practice due to legal/ethical concerns.