Loading field map
Backgammon's opening moves present a unique analytical challenge. Unlike chess, where the first few moves can be memorized from a deterministic tree, backgammon's opening rolls are probabilistic: each of the 21 possible first rolls can be played in multiple ways, and the opponent's reply introduces another layer of chance. For decades, players relied on accumulated experience and general principles. The history of opening theory is the story of how that human authority was gradually displaced—and then supplemented—by systematic analysis, machine learning, statistical verification, and finally a community-maintained synthesis that integrates all of these.
Before computers, opening knowledge was passed down through expert consensus. The Pre-Bot Expert Consensus framework (roughly 1960–1993) consisted of a loose collection of heuristics, positional rules, and published recommendations from top players. Authors like Paul Magriel, in his landmark 1976 book Backgammon, codified many of these principles: make points, hit blots, escape runners, and avoid leaving direct shots. But the consensus was far from uniform. Different schools of thought coexisted, and the same opening roll often had multiple recommended plays depending on which expert you consulted.
Within this broad consensus, a narrower methodological school emerged in the mid-1970s: the Five-Point Slotting Doctrine. This doctrine prescribed a specific approach to the opening: slot the five-point (that is, place a checker on the five-point in your home board) as aggressively as possible, even at the cost of leaving blots. The idea was that securing the five-point early gave such a strong positional advantage that the risk was worth it. For nearly two decades, this doctrine shaped how many strong players handled rolls like 2-1, 3-1, and 4-1. Yet it was never universally accepted; critics argued that the doctrine was too rigid and that other positional considerations—such as building a prime or escaping back checkers—were sometimes more important.
The first major break from the human-authority era came with Opening Reply Theory (1976–Present). Rather than focusing only on the first roll, this framework systematically analyzed the second roll: given any opening move by the opponent, what is the best reply? The key innovation was methodological. Instead of relying on general principles, Opening Reply Theory organized the problem into a 21×21 matrix—one row for each possible opening roll, one column for each possible reply roll—and attempted to fill each cell with a recommended play based on positional reasoning and selective analysis.
This framework did not reject the Pre-Bot Expert Consensus wholesale; it absorbed much of its positional vocabulary. But it narrowed the focus from broad heuristics to a position-by-position catalog. Jeremy Bagai's 1998 book Bagai's Replies was a landmark in this tradition, offering a comprehensive set of recommendations for the second roll. Opening Reply Theory remains active today, especially as a pedagogical tool: it teaches players to think about the opening as a sequence of interdependent decisions rather than isolated first-roll choices. However, its recommendations have been repeatedly revised as computational methods revealed that many of its human-derived judgments were inaccurate.
The arrival of neural-network bots in the early 1990s transformed opening theory from a human-centered craft into a data-driven science. Neural-Net Bot Opening Theory (1993–Present) began with TD-Gammon, a program developed by Gerald Tesauro at IBM. TD-Gammon learned to evaluate backgammon positions by playing against itself millions of times, using temporal-difference learning. Its evaluations of opening moves often contradicted the established human consensus. For example, TD-Gammon recommended slotting the five-point less aggressively than the Five-Point Slotting Doctrine prescribed, and it favored plays that human experts had considered inferior.
This framework did not simply add new data to the old consensus; it overturned the authority of human judgment. For the first time, players had access to an objective evaluator that could assess any position—including opening positions—with a consistency no human could match. The neural net did not explain its reasoning in human terms, but its recommendations were demonstrably stronger than those of any living player. This created a tension that persists today: should players trust the bot even when its recommendations contradict established positional principles?
Neural-net evaluation was powerful, but it had a limitation: it produced a single numerical estimate of equity, not a confidence interval. Computer Rollout Opening Theory (1997–Present) addressed this by using rollouts—playing out a position many thousands of times with random dice sequences—to produce statistically robust equity estimates. Rollouts could verify or challenge the evaluations of neural nets, and they often did. For instance, rollout-based analysis showed that some plays favored by early neural nets were actually inferior when the full range of future dice sequences was considered.
Rollout-based analysis coexists with neural-net evaluation as a complementary method. The neural net provides a fast, approximate evaluation; the rollout provides a slower but more accurate verification. Together, they form the backbone of modern opening analysis. The Score-Specific Opening Theory framework (2006–Present) extended this computational infrastructure by adding a new variable: the match score. In tournament backgammon, the optimal opening play can change depending on whether you are leading, trailing, or playing for a gammon. Score-Specific Opening Theory uses neural-net and rollout tools to compute opening recommendations for every possible match score, not just for money play. This framework is a direct application of computational methods to a contextual problem that human experts had only vaguely addressed.
The most recent framework, the XG Community Opening-Book Consensus (2011–Present), is not a new analytical method but a collaborative verification and synthesis model. It is built around eXtreme Gammon (XG), a commercial neural-net bot that includes a community-maintained opening book. The opening book is continuously updated as players submit rollout results for specific positions, and the community votes on which plays to include. This framework integrates all the earlier methods: neural-net evaluation for initial estimates, rollouts for verification, score-specific adjustments for tournament play, and human oversight for edge cases where the bot's recommendations are ambiguous or contradictory.
Today, the leading frameworks divide the labor of opening theory as follows. Neural-Net Bot Opening Theory provides the core evaluative engine; Computer Rollout Opening Theory supplies statistical rigor; Score-Specific Opening Theory adapts the results to match contexts; and the XG Community Opening-Book Consensus curates and disseminates the findings. Opening Reply Theory remains useful as a teaching tool and as a framework for organizing the second-roll problem, though its specific recommendations are now routinely overridden by computational analysis. The Pre-Bot Expert Consensus and the Five-Point Slotting Doctrine are largely of historical interest, though some of their positional insights survive in modified form.
There is broad agreement that neural-net evaluation combined with rollout verification produces the most reliable opening recommendations. No serious analyst today would rely on human judgment alone. There is also agreement that score-specific adjustments matter: the same opening play can be correct at one match score and incorrect at another. The main disagreement concerns the role of human oversight. Some argue that the community consensus model (XG) is too conservative, because it can be slow to incorporate new findings that contradict established recommendations. Others argue that the bots themselves are not infallible—they can make systematic errors in rare or unusual positions—and that human judgment remains essential for interpreting their output. This tension between computational authority and human community governance is the defining debate of contemporary opening theory.
The arc of backgammon opening theory is a movement from folk rules to a computationally rigorous, score-sensitive, and collaboratively verified science. Each framework addressed a limitation of its predecessors: the Pre-Bot Expert Consensus lacked systematic verification; the Five-Point Slotting Doctrine was too narrow; Opening Reply Theory was human-centered and error-prone; Neural-Net Bot Opening Theory lacked statistical confidence intervals; Computer Rollout Opening Theory was slow and context-insensitive; and Score-Specific Opening Theory required a synthesis mechanism. The XG Community Opening-Book Consensus is the current endpoint of this evolution, but it is not the final word. As bots become stronger and rollout methods faster, the community will continue to refine its understanding of the opening—and the debate over human versus engine authority will likely persist.