Chibi was an early computer program for playing checkers (draughts), developed by Arthur Samuel at IBM in the 1950s. It is widely regarded as one of the first successful demonstrations of artificial intelligence and machine learning, predating modern deep learning and neural network approaches. Samuel's work on Chibi laid foundational concepts for game-playing AI and adaptive algorithms, influencing subsequent research in the field.
The program was designed to learn from experience rather than relying solely on pre-programmed strategies. Samuel implemented a scoring function based on board features such as piece count, king advancement, and positional control, which could be adjusted through a process of self-play and comparison against stored evaluations. This allowed Chibi to improve its play over time, a notable early example of reinforcement learning principles.
Development and Historical Context
Arthur Samuel began working on Chibi in 1952 while at IBM, using the IBM 701 computer - one of the first commercially available scientific computers. The program's name 'Chibi' is often noted as a playful reference, though its exact origin is not well documented. Samuel's research was part of a broader effort in the 1950s to explore whether machines could exhibit intelligent behavior, alongside other pioneering projects like the Logic Theorist and early chess programs.
Chibi was developed during a period when computing resources were extremely limited. The IBM 701 had only about 2,000 words of memory, which constrained the complexity of the evaluation function and search depth. Samuel circumvented these limitations by using a minimax search with alpha-beta pruning, a technique that reduced the number of positions examined, and by storing learned weights in a separate memory bank.
Learning Mechanism
Chibi's learning process was based on a technique Samuel called 'rote learning' and 'generalization learning'. In rote learning, the program stored the value of each board position it encountered during play, using these stored values to guide future decisions. In generalization learning, it adjusted the weights of a linear evaluation function based on differences between predicted and actual outcomes, a method analogous to modern gradient descent but implemented without formal calculus.
Samuel also introduced a form of self-play where Chibi played against itself, with one version using a current evaluation function and another using a modified version. The program compared their moves and updated the weights to favor the more successful strategy. This approach was a precursor to techniques used in later AI systems, such as residual networks and reinforcement learning (though the latter term was not in use then).
Performance and Impact
By the late 1950s, Chibi had achieved a level of play that could compete with amateur human players, and in some demonstrations, it defeated a former Connecticut checkers champion. Samuel's work was featured in a 1956 IBM press release and later in a 1959 article in the IBM Journal of Research and Development, which became a classic reference in AI literature.
The program's success helped establish the viability of machine learning in game playing, influencing subsequent work on chess computers and other game AI. Samuel's emphasis on learning from experience rather than explicit programming was a key departure from earlier rule-based systems, and it anticipated the importance of data-driven approaches in modern AI.
Legacy and Modern Relevance
Chibi is often cited in histories of AI as a milestone, though it is less well-known than later programs like IBM's Deep Blue. Its techniques - heuristic evaluation, minimax search, and adaptive weight adjustment - remain relevant in contemporary game AI and reinforcement learning research. The concept of self-play, which Chibi pioneered, is now a cornerstone of systems like AlphaGo and modern large language models that use self-generated data for training.
Samuel's work also highlighted the importance of computational efficiency, a concern that persists in modern AI with techniques like model pruning and data augmentation. While Chibi itself is no longer operational, its principles are embedded in many current AI systems, and its historical significance is recognized in academic courses on artificial intelligence.
Technical Details
Chibi's evaluation function considered several features: the number of pieces, the number of kings, the number of pieces on the back row, and a measure of mobility (the number of possible moves). Each feature was assigned a weight, and the total score was computed as a weighted sum. The program used a minimax search with a depth limit, typically looking ahead 3 to 5 moves, and applied alpha-beta pruning to reduce the search tree.
Samuel also implemented a 'signature' system to identify board positions that were similar, allowing the program to generalize from one position to another. This was an early form of feature extraction, though it was not based on neural networks. The learning rate was controlled by a parameter that determined how much the weights were adjusted after each game, a concept similar to modern learning rate schedules.
Chibi ran on the IBM 701 and later on the IBM 704, which had more memory and speed. Samuel continued to refine the program until the early 1960s, when he moved to other projects. The source code and documentation are preserved in some archives, and the program has been recreated in emulators for educational purposes.