When does planning beat learning?
mediumAnswer
- (1) Compact known dynamics (games, physics).
- (2) Long horizons where credit assignment is hard for learning.
- (3) Rare edge cases (planning can reason zero-shot; learning needs data).
- (4) Test time compute available.
- AlphaZero + MuZero show planning + learning outperform either alone.
- Modern LLM reasoning (o1, DeepSeek-R1) is analogous: test-time 'thinking' as planning augments learned priors.
Check yourself — multiple choice
- Never
- Compact known dynamics / long horizons / rare edge cases / test-time compute — planning + learning > either alone (AlphaZero, MuZero); LLM reasoning is analogous
- Random
- Always
Planning wins: known dynamics + long horizons + rare cases + test-time compute.
#planning#theory
Practise Reinforcement Learning
214 interview questions in this topic.