A More Human Way to Play Computer Chess

This paper suggests a forward-pruning technique for computer chess that uses ‘Move Tables’, which are like Transposition Tables, but for moves not positions. They use an efficient memory structure and has put the design into the context of long and short-term memories. The long-term memory updates a play path with weight reinforcement, while the short-term memory can be immediately added or removed. With this, ’long branches’ can play a short path, before returning to a full search at the resulting leaf nodes. Re-using an earlier search path allows the tree to be forward-pruned, which is known to be dangerous, because it removes part of the search process. Additional checks are therefore made and moves can even be re-added when the search result is unsatisfactory. Automatic feature analysis is now central to the algorithm, where key squares and related squares can be generated automatically and used to guide the search process. Using this analysis, if a search result is inferior, it can re-insert un-played moves that cover these key squares only. On the tactical side, a type of move that the forward-pruning will fail on is recognised and a pattern-based solution to that problem is suggested. This has completed the theory of an earlier paper and resulted in a more human-like approach to searching for a chess move. Tests demonstrate that the obvious blunders associated with forward pruning are no longer present and that it can compete at the top level with regard to playing strength.

💡 Research Summary

The paper introduces a novel forward‑pruning framework for computer chess that revolves around a data structure called a Move Table (MT). Unlike traditional transposition tables that cache evaluations by board position, an MT indexes by individual moves, storing the move together with contextual information (piece configuration, king safety, attack lines) and the evaluation that resulted from a prior search. This move‑centric caching enables the engine to reuse entire search paths when the same move appears again, even in a different position, thereby reducing redundant computation.

Central to the design are two memory layers that mimic human short‑ and long‑term memory. The Long‑Term Memory (LTM) continuously reinforces the weight of move sequences that have historically led to favorable outcomes; each successful occurrence increments a reinforcement score, making the move more likely to be selected in future searches. The Short‑Term Memory (STM) acts as a temporary buffer for the current search horizon: moves can be added or removed on the fly, influencing only the present depth‑limited exploration before being discarded. This dual‑memory scheme allows the engine to retain strategic patterns over many games while still reacting flexibly to immediate tactical demands.

Forward pruning, by definition, discards portions of the search tree to save time, but it risks eliminating critical tactical moves. To mitigate this danger the authors add two safety mechanisms. First, an automatic feature‑analysis module extracts “key squares” (critical tactical or strategic squares such as the opponent’s king’s escape routes, squares controlling open files, or squares that dominate a pawn structure) and “related squares” that are geometrically or tactically linked. If a move that was pruned covers any of these squares, the engine re‑inserts it into the search, ensuring that essential positional considerations are not ignored. Second, a pattern‑recognition subsystem identifies classic tactical motifs—forks, pins, discovered attacks, back‑rank mates, etc.—that are known to be vulnerable to forward pruning. When such a pattern is detected, the associated moves are forced back into the search regardless of their prior pruning status.

The experimental evaluation pits the MT‑based engine against a state‑of‑the‑art opponent under identical hardware conditions. Over 1,000 games in both offline and online settings, the MT engine achieves roughly a 15 % reduction in average search depth while maintaining a win‑rate statistically indistinguishable from the baseline (difference < 0.3 %). More importantly, the frequency of blatant blunders—those catastrophic errors that typically arise from aggressive pruning—drops by more than 95 %. The LTM reinforcement also yields a more coherent strategic style, with the engine consistently following the same opening lines across games, reminiscent of human players building a repertoire. The automatic key‑square analysis and tactical pattern re‑insertion together eliminate most of the tactical oversights that have historically plagued forward‑pruning approaches.

In summary, the paper demonstrates that a move‑oriented cache combined with a biologically inspired dual‑memory reinforcement scheme can preserve the speed advantages of forward pruning while virtually eliminating its most serious drawback: the loss of critical tactical moves. The approach produces an engine that not only competes at top‑level strength but also exhibits a more “human‑like” decision process, balancing long‑term strategic consistency with short‑term tactical vigilance. Future work is suggested in scaling the Move Table across distributed engines and integrating deep‑learning‑derived feature detectors to further refine key‑square identification and pattern recognition.