Product Sheet : SumoBots Project
In order for SumoBots to play and react against human players in an autonoumous manner, SumoBot uses artificial intelligence that is managed by a parameterized expert system. The evolution and learning aspects of SumoBot's artificial intelligence represent a complex problem that requires optimal semantic parameterization of this expert system.
In fact, it is very difficult to know ahead of time what the best game tactics will be. This is due partially to the fact that there can exist several different effective tactics - each with its own strong and weak points, and also to the fact that combining several basic tactics can actually lead to the creation of new tactics that are even more efficient.
For these reasons, and because the game structure and number of SumoBots makes it possible, we have chosen to have the respective intelligence of individual SumoBots evolve by using a genetic algorithm that is completely distributed across the population of existing SumoBots.
The use of genetic algorithms, and more generally of biological metaphors for controlling processes, is at least partially rooted in the difficulties encountered by traditional command theory with respect to environments that are complex, uncertain, and variable over time. This is exactly the case we encountered when defining the roboticized behavior of SumoBots. These difficulties lie mainly in the fact that even adaptive command theory requires fairly precise and thorough knowledge of the processes to be controlled (among other things, one needs basic knowledge of the process structure and possible perturbations).
However, it is not easy to come up with a model that defines the player who facing a SumoBot, especially if that player is a child. On top of that, genetic algorithms, much in the same way as other computerized artificial intelligence technologies, such as artificial neuron networks, are characterized by the ability to reconfigure themselves in an incremental and robust manner, using principles often similar to self-organization, with respect to an unknown environment.
These biological metaphors do not require as many restrictive hypotheses or as much information about processes. They therefore represent the approach that is best suited for adaptive control of the roboticized SumoBot's game processes. This approach is also reinforced by the parallel nature of the overall algorithm and by recent developments in parallel processing.
As mentioned above, SumoBot's artificial intelligence is based on a genetic algorithm specially developed for the robot's tiny microprocessor. The algorithm controls the creation and evolution of strategies based on its success or failure in combat.
For example, strategies are considered as sequences of limited actions defined in the SumoBot in regard of the situation perceived by the sensors. Their usage depends on the position, the orientation and the speed of the two SumoBots on the dohyo, and is determined by the artificial intelligence on the fly.
The results of each strategy are then analyzed so that the SumoBot can learn from its experience and become more powerful.
In addition to this self learning behaviour, SumoBots can exchange segments of DNA after each battle. The winner robot gain the ability to reproduce a segment of its DNA on the loser genes, at the same position. During the exchange, the segment of DNA has a certain probability to be altered.
Winning robots can evolve also by occasionally performing their own genetic cross-over: once in approximately twenty combats, the loser randomly changes a parameter in its chromosome structure. This allows the best robots to continue getting better even if they are not always successful in the competition.
The global algorithm is completely distributed across the entire population. That is, each SumoBot (each individual) participates - through its personal evolution, in the global convergence of the algorithm that is actually operating on the entire population of potentially victorious SumoBots (individuals or chromosomes).
An online game has been developed to illustrate the concept of SumoBots.