The paper demonstrates that we can evolve basic behaviors and adapt to the environment using CHC, a non-traditional genetic algorithm. Injecting selected solutions stored in a long term memory and corresponding to these basic behaviors into the GA's initial population allows us to quickly and successfully design control strategies for a robot navigating in a complex office environment. The experimental results are promising and the simulated robot is faster and accomplishes more of the task than the robot designed by a randomly initialized GA. We are currently investigating parallelization of the code to handle a larger population size in a reasonable amount of time. This will allow us to handle more complex environments. We are also planning to transfer the evolved circuits to a real mobile robot, thus testing our work on physical hardware with all its concomitant problems. We will be investigating the effect of noise on performance - circuits evolved in the presence of noise may be more robust and better able to handle the noise inherent in a real mobile robot operating in a complex environment.
We have only reported on non-randomly initializing genetic algorithms in this paper. However, the concept is extendable to other population based searches like evolutionary programming, evolution strategies, and genetic programming. In addition, there is no reason why injection of individuals should only take place at initialization - we can inject individuals during the course of GA's run. We believe that investigating the combination of population-based search algorithms with a long term memory promises to be a fruitful area of future research. The hope is that as the number of problems solved by the combined system grows, the time taken to solve a new problem shrinks.