1959: a checkers program that learned by playing itself
You play navy and move first. The program scores positions with a weighted sum of six board features – using random weights.
What the program values (its weights)
The 1959 original
- Arthur L. Samuel, IBM. Ran on the IBM 704, usually looked three moves ahead in the opening, about 30 seconds per move.
- Its scoring polynomial had 38 terms, 16 in use at any one time. Here: 6 features, all in use.
- Goal: make a computer “behave in a way which, if done by human beings or animals, would be described as involving the process of learning”.
Alpha plays Beta. Only the weights change.
Alpha won
Beta won
draws
shake-ups (3 losses in a row)
Live self-play game
waiting
Draw after 80 moves (40 each) without a capture.
The weights while it learns
Look-ahead (minimax)
How Alpha learns – learning by generalisation (simplified)
- Alpha adjusts its weights during the game; Beta keeps its weights fixed for the whole game.
- After each Alpha move: delta = look-ahead score − quick score of the position. The minimax look-ahead sees further than the quick weighted sum.
- Each weight is nudged in proportion to delta × that feature's value (normalised, small learning rate, clipped).
- If Alpha wins, Beta receives Alpha's weights. After 3 losses in a row Alpha gets a random shake-up (Samuel's “black marks”, simplified).
- Rote learning runs alongside: positions are remembered with their look-ahead scores and reused in later searches.
P is measured every 5 self-play games: a 16-game match against the untrained start (8 fixed openings × both colours, both sides 2 moves ahead, no memory). Draws count ½.
Now play the trained program
Same board, same rules, same look-ahead. Compare this game with your game in step 1.
You vs untrained (step 1)
won–lost–draw
You vs trained (after 0 games)
won–lost–draw
What the program values now
What changed?
Only the numbers (weights). The rules, the code and the search did not change – the experience did.
Myth check: how good was it really?
Samuel's own judgement: rote learning gave “a rather better-than-average novice, but definitely not … an expert”; the generalisation program “a better-than-average player”.
The often-told 1962 win against Robert Nealey is overstated: Nealey was a blind checkers player from Connecticut whose tournament results never justified IBM's description as “one of the nation's foremost players” (Schaeffer, 1997).
Checkers was finally solved in 2007: with perfect play by both sides it is a draw (Schaeffer et al., 2007).
