Learning in Problem Solving

Learning in problem solving means a system improves its own problem-solving performance by reflecting on its own past attempts — successes and failures — rather than by being told new facts or rules from outside.

Key Mechanisms

Learning from Success: Macro-Operators

If a particular sequence of operators repeatedly proves useful for solving a class of sub-problems, the system can package that sequence into a single new macro-operator, so future problems of the same shape are solved in one step instead of being re-derived from scratch.

Example (Blocks World): if "pick up block, move to target, place on top" repeatedly appears as a useful sub-sequence across many block-stacking problems, this becomes a reusable macro-move rather than being re-planned every time.

Learning from Failure: Failure-Driven Learning

When a solution attempt fails, the system analyses why, and stores that reason as a new constraint or heuristic to avoid repeating the same mistake — this is closely related to explanation-based learning (a following lesson), but focused specifically on problem-solving traces rather than a single labelled example.

Learning Better Heuristics

A problem solver can also observe, across many solved problems, which heuristic evaluation function actually correlated with shorter/cheaper solutions, and adjust its own heuristic accordingly — effectively tuning its own search guidance from experience.

Comparing to Other Forms of Learning

Why It Matters

This form of learning is what lets a problem solver become genuinely more skilled at a class of problems over time — not just faster on exact repeats (rote learning), and not dependent on a human constantly supplying advice — the system is, in effect, coaching itself using its own history.