The AI landscape — one map¶
Every AI approach you meet on this platform answers the same question differently: where do the rules come from?
| Approach | Rules come from | You provide | Breaks when |
|---|---|---|---|
| Rule-based | you write them | logic & thresholds | the world surprises you |
| Supervised learning | examples you label | labeled data | data doesn't cover a case |
| Reinforcement learning | trial & error | a reward function | the reward is mis-designed |
| LLM agents | pretrained knowledge + tools | instructions & tools | the task exceeds its tools |
Rule-based¶
Transparent and instant — you can read exactly why the robot turned. But every exception becomes another rule, and rules multiply until they fight each other. You felt this in Line Tracing.
Machine learning (the next three)¶
Instead of writing rules, you provide experience and the machine extracts the rules itself — as millions of numbers (weights) rather than readable statements. What kind of experience distinguishes the branches:
- Labeled examples → supervised learning
- A reward signal → reinforcement learning
- Almost everything ever written, then steered by your text → LLMs & agents
Why learn them all on one robot¶
Methods are trade-offs, not upgrades. Rule-based wins when rules are few and known; supervised wins with good data; reinforcement wins when you can simulate; agents win on open-ended tasks. Driving the same track with each is the fastest way to feel those trade-offs.
