Arthur Samuel

Arthur Samuel

Electrical engineer and machine-learning researcher · 1901–1990 · United States

A game-playing program can search farther, remember earlier results, or change the rule it uses to score positions. These are different mechanisms. Samuel's checkers experiments make it possible to study their interaction without claiming that a program escapes the algorithm that implements its learning.

Correcting the career record

Arthur Samuel was born in Emporia, Kansas, in 1901 and died on 29 July 1990. He received an MIT master's degree in electrical engineering in 1926, not an earned doctorate. After teaching at MIT, he joined Bell Telephone Laboratories in 1928. His Illinois professorship lasted from 1946 until his move to IBM in 1949, not from 1926 to 1946. He joined Stanford in 1966. Stanford memorial

His 1959 paper, Some Studies in Machine Learning Using the Game of Checkers, documented experiments with two learning procedures. The title is direct evidence of his use of “machine learning”; it is not by itself proof that no one had ever used the words before. Broad claims that every later learning system descends from this one paper are unnecessary. Primary paper

IBM dates the publicized win against Robert Nealey to 1962, not 1961, and notes a mixed subsequent record. One win does not establish world-champion strength, stable improvement, or mastery of every position. IBM retrospective

Separate search from learning

A bounded game-tree search examines possible continuations. At a maximizing node, choose the largest child value; at an opponent's minimizing node, choose the smallest. If the leaves are terminal positions and scores faithfully describe outcomes, minimax reasons about the represented game. If the leaves are cutoff positions scored heuristically, its answer is only as reliable as that evaluation and horizon.

Rote learning stores particular positions and associated information. Generalization adjusts a representation that also applies to other positions, for example coefficients of a feature-based evaluation function. A remembered search value need not be an exact game-theoretic outcome; it may depend on depth, perspective, and the evaluation version. Samuel explicitly discussed limitations and interactions between his learning methods. Search and learning in the 1959 paper

Neither method removes programming. Rules, move generation, search management, features, and update procedures still require specification. Learned parameters change within that programmed process. The scientific question is whether the change improves measured performance on an appropriate task.

Worked mechanism: search can correct a tempting move

This original toy tree uses scores from one fixed player's perspective. The player chooses A or B, and an adversary then chooses a reply:

Root move	Reply scores	Backed-up minimum
A	8, -3	-3
B	2, 1	1

A greedy choice based on the best visible leaf would choose A because 8 exceeds 2. Minimax chooses B: an adversary can force -3 after A but only 1 after B. Averaging the replies would answer a different question, requiring a probability model for the opponent rather than worst-case play.

Now suppose the current position has feature vector φ=(1,2), with weights w=(0.5,0.25). Its linear estimate is V=w·φ=1. A deeper search returns target value y=-1. For an illustrative gradient update on squared error, define

error = y - V = -2
w_next = w + α error φ, with α=0.1
       = (0.5,0.25) + (-0.2,-0.4)
       = (0.3,-0.15).

The new estimate at the same feature vector is 0.3+2(-0.15)=0. Its error magnitude fell from 2 to 1. This shows one local update, not Samuel's exact historical coefficient algorithm. It is a modern simplified model illustrating how search-derived information could train an evaluator.

At another position with features (1,-2), the old prediction was zero and the new one is 0.3+0.3=0.6. Learning generalized to a position not used in the update. Whether that change helps is unknown without a trustworthy target or outcome for the other position. Generalization can spread errors as well as useful information.

Evaluate an improvement claim

A self-play win rate can change because both opponents change. Preserve fixed baselines and separate training experience from evaluation games. Record search budget, opening distribution, color assignments, random seeds where relevant, and the treatment of draws. Report outcomes rather than only an impressive anecdote.

For example, over 40 held-out games a candidate scores 18 wins, 12 draws, and 10 losses. With one point per win and half per draw, its score rate is (18+6)/40=0.60, while its win rate is 18/40=0.45. These statistics answer different questions. Neither proves superiority across all openings or budgets. Games may also be dependent because they share openings or training histories.

Caching creates another validation issue. A stored value computed by evaluator version v1 at depth 3 should not silently be treated as an exact value under version v2 at depth 8. Metadata can preserve provenance: side to move, search depth, bound type, evaluation version, and whether the value is terminally exact. This is an original modern design lesson, not a claim that Samuel used this precise record format.

Exercises

1. Add root move C with opponent replies 4,0,5. Which move does minimax choose among A, B, and C? Would a uniformly random opponent produce the same ranking? Calculate the expected values under that separate assumption.
2. Repeat the weight update with α=0.3 instead of 0.1. Compute the new prediction at (1,2) and compare absolute error with the original. Derive the factor by which the error changes for this one-example linear update.
3. A candidate wins 11, draws 8, and loses 5 of 24 held-out games. Calculate win rate and half-point score rate. Describe two changes to the evaluation protocol that would strengthen a claim of learning, and one conclusion these numbers cannot justify.

Compare your reasoning with the worked solutions

1. C has minimum zero, so minimax still chooses B with value 1, ahead of C's zero and A's -3. For a uniformly random opponent within each branch, A averages (8-3)/2=2.5, B averages (2+1)/2=1.5, and C averages (4+0+5)/3=3. Expected-value choice would favor C. The difference follows from a changed opponent model, not an arithmetic contradiction.
2. The increment is 0.3×(-2)×(1,2)=(-0.6,-1.2), giving weights (-0.1,-0.95). The new prediction is -2, overshooting target -1; the absolute error is 1, still below the original 2. In general V_next=V+α(y-V)||φ||², so error_next=(1-α||φ||²)error. Here ||φ||²=5, producing factor -0.5. A sufficiently large step can increase error; local improvement is not automatic.
3. Win rate is 11/24≈45.8%. Score rate is (11+4)/24=62.5%. Use fixed opponents and balanced colors, and evaluate on openings excluded from training; recording equal computation budgets also matters. These results do not establish optimal play, universal superiority, or statistical independence of the games. A numerical score is evidence under a protocol, not proof of general intelligence.