This combination thesis within Computer Science and Communication utilizes the board game Checkers as a case to reveal the basic elements for the generally acknowledged success using artifi-cial intelligence in board games and to which extent these experiences and methods can be general-ized to include aspects of human-like intelligence. Artificial intelligence at this level is referred to as strong artificial intelligence. The theory behind artificial intelligence for board games is outlined, in-cluding context issues such as the frame problem, Hamlet’s problem, the prediction problem and the commonsense-knowledge problem, where in particular the frame problem poses difficulties for de-veloping artificial intelligence beyond rule governed cases. A so called MinMax algorithm implemen-tation has been optimized by use of methods to reduce calculation time known as Alpha-Beta prun-ing, Move sorting and Hash table. MinMax algorithm has, due to its algorithmic approach strategic weaknesses leading to unsuitable behavior. This has led to experiments with a paradigm hybrid where the evaluation function has been replaced by a neural network in an attempt to create a high level stra-tegic artificial player and to investigate its potential for approaching strong artificial intelligence. The neural network has been implemented using temporal-difference method, by which experience from earlier draws is part of incremental learning. Various types of training passes for neural net-works and tests of several artificial players at strategically different levels have been conducted. As input methods for neural networks both board-mapping and feature-mapping has been implemented and tested. With a phenomenological approach test persons have been observed while playing Check-ers. The objective of the empirical study was to unravel how human players act and interact while playing Checkers and how this would relate to abilities of an artificial player as described above. We discuss views and statements from selected, renowned artificial intelligence researchers on the phi-losophical perspectives on strong artificial intelligence. Based on the case study and the discussion on philosophical perspectives we conclude that current methods used in artificial players for Checkers does not have the potential for including intelligent behavior and that current level of artificial intelligence for board games is far from strong artificial intelligence. From the artificial player implementations and tests we conclude that the combination of the two paradigms MinMax algorithm and neural networks shows good performance and adds to the algorithm ability to adapt to opponent behavior and to learn from own faults. Thus neural networks do compete favorably with static evaluation functions and the combination is considered a suitable ex-tension of the MinMax algorithm. However, the combination cannot bring the artificial player of Checkers closer to strong artificial intelligence. We can conclude that an artificial player at the level investigated cannot participate in the psychological game which plays an important role when humans play. Along with the philosophical discussion we conclude that an artificial player can be trained to play Checkers at high level but not intelligent behavior as board games is strictly rule governed with-out any unforeseen events. Artificial intelligence for board games cannot be used directly to create intelligent behavior. The artificial player does not think, nor is he understanding or being conscious, when selecting his move.
|Uddannelser||Datalogi, (Bachelor/kandidatuddannelse) KandidatKommunikation, (Bachelor/kandidatuddannelse) Kandidat|
|Udgivelsesdato||30 maj 2007|
- Hamlets problem
- kunstig intelligens
- neurale netværk