An exposed checker was once the confident choice in a familiar backgammon opening. After a roll such as 2–1, experts would “slot” onto a valuable point, risking a hit to build a stronger position. IBM researcher Gerald Tesauro’s 1995 account described roughly 30 years of expert preference for that approach.[1]
TD-Gammon preferred to split the rear checkers. Tesauro reported that Bill Robertie publicised its analysis in 1992; top players experimented, and the alternative spread. His account describes early-1990s strategy, rather than advice about today’s best opening.[1]
Learning without a catalogue of correct moves
TD-Gammon acquired its judgement through games against itself. The network evaluated board positions, used those evaluations to choose moves and adjusted its predictions as play unfolded. The 1993 technical report describes training sequences that ended in a game result, supplying the outcome from which earlier judgements could learn.[2]
The progression matters. Tesauro’s 1992 paper distinguished a system given a raw encoding of the board from one also supplied with human-designed features. The former reached an intermediate standard; the added features helped it approach expert play. Self-teaching described how it learned strategy, not an absence of human work in constructing the program.[3]
By the 1993 report, version 2.1 had trained on 1.5 million games and finished a 40-game test against Robertie just one point behind. That is a reported points result from a short contest, not proof that it would always beat him. It does show why its unusual suggestions deserved serious investigation.[2]
How tomorrow’s estimate teaches today’s
The learning idea came from temporal differences, developed in Richard Sutton’s 1988 paper. A predictor can learn from the change between successive estimates, rather than waiting for a final answer every time it adjusts. Later information becomes a teaching signal for an earlier judgement.[4]
Consider a deliberately simplified illustration. A system rates a position at 0.40, then rates its successor at 0.60 from the same player’s perspective. With no immediate reward or discounting, moving a tenth of the way towards that later estimate gives 0.40 + 0.10 × (0.60 − 0.40) = 0.42. These are invented teaching values, not measurements from TD-Gammon.
A small correction is the point of the example. The earlier prediction need not become an unquestioning copy of the later one. Repeated experience supplies further corrections. A completed game eventually supplies an outcome rather than another estimate; the full learning algorithm handles a sequence of such updates.[4]
From opponent to instrument
Its opening rollouts compared alternative moves through many simulated continuations. Players could examine two choices through repeated trials instead of treating one win as decisive.[1]
The striking historical contribution was this change in the conversation. An opening passed down through expert practice became a question that a learning system could help reopen. The human task expanded from choosing a move to deciding which inherited assumptions were ready to be tested.
Sources & further reading
- Temporal Difference Learning and TD-Gammon Gerald Tesauro, Communications of the ACM; reproduced by Backgammon Galore · Accessed 19 Oct 2026
- TD-Gammon, A Self-Teaching Backgammon Program, Achieves Master-Level Play Gerald Tesauro, AAAI · Accessed 19 Oct 2026
- Temporal Difference Learning of Backgammon Strategy Gerald Tesauro, IBM Research / ICML 1992 · 1 Jul 1992 · Accessed 19 Oct 2026
- Learning to Predict by the Methods of Temporal Differences Richard S. Sutton, Machine Learning · Accessed 19 Oct 2026