Blondie24 Reconsidered
I discovered David Fogel's "Blondie24: Playing at the Edge of AI" in 2006 while killing time at the library waiting for my daughter. It is a very engaging story of his effort, with colleague Kumar Chellapilla, to evolve an American Checkers (English Draughts) "player", with no human expertise provided, and then to evaluate that player by competing online against (presumably) people on zone.com (now zone.msn.com).
Dr. Fogel covers the history of AI in playing chess and checkers, and provides background on evolutionary algorithms and artificial neural networks. He then covers the three network topologies they used, and the training algorithm, the key feature of which is that they created a random of population of players (really, evaluation functions), and had those players, and their offspring, compete against each other in tens of thousands of checkers games across hundreds of generations, keeping and evolving the most successful players..
One of the challenges they faced was limited computing resources; I think they were doing this work around the year 2000, and that the machine they had available to (mostly) dedicate to this task was a PC with a 400MHz Pentium processor. For their final network topology, in which their were more than 5000 weights to be adjusted, it took 6 months to run through 800 generations. As a result, they were not able to explore the many lines of investigation that spring to mind.
I'm interested in following some of those lines, and I have the luxury of 10 more years of CPU development. To provide a baseline, I've been working to reproduce their work, based on the information in Dr. Fogel's book, and the several papers they published together on the topic. I'll be writing about my efforts in posts to follow this.
Dr. Fogel covers the history of AI in playing chess and checkers, and provides background on evolutionary algorithms and artificial neural networks. He then covers the three network topologies they used, and the training algorithm, the key feature of which is that they created a random of population of players (really, evaluation functions), and had those players, and their offspring, compete against each other in tens of thousands of checkers games across hundreds of generations, keeping and evolving the most successful players..
One of the challenges they faced was limited computing resources; I think they were doing this work around the year 2000, and that the machine they had available to (mostly) dedicate to this task was a PC with a 400MHz Pentium processor. For their final network topology, in which their were more than 5000 weights to be adjusted, it took 6 months to run through 800 generations. As a result, they were not able to explore the many lines of investigation that spring to mind.
I'm interested in following some of those lines, and I have the luxury of 10 more years of CPU development. To provide a baseline, I've been working to reproduce their work, based on the information in Dr. Fogel's book, and the several papers they published together on the topic. I'll be writing about my efforts in posts to follow this.
Comments
Post a Comment