Asynchronous Migration in Parallel Genetic Programming

Genetic Programming (GP) was used to generate robot control programs for an obstacle avoidance task [1]. The task was to control an autonomous mobile robot from a starting point to a target point in a simulated environment. The environment was filled with the obstacles which had several geometrical shapes. In order to improve the robustness of the program, each program was evaluated under many environments. As a result, the substantial processing time was required to evaluate the fitness of the population of the robot programs. To reduce the processing time, this present study introduced a parallel implementation. In applying the parallel approach to the algorithm program by using a conventional coarse-grained model, the result achieved only linear speedup since the amount of work was fixed – the algorithm was terminated when it reached the maximum generation. Hence, the parallel algorithm did not exploit the probabilistic advantage that the answer may be obtained before the maximum generation. We tried in this present study another method to further improve the speedup by dividing the environments among the processing nodes. After a specific number of generations, every subpopulation was migrated between processors using a fully connected topology. The parallel algorithm was implemented on the dedicated cluster of PC workstations with 350 MHz Pentium II processors, each with 32 Mb of RAM, and running Linux as an operating system. These machines were connected via 10 Mbs ethernet cabling. We extended the program used in [1] to run under the clustered computer by using MPI as a message passing library. In the first stage of the implementation, the migration was synchronized. The synchronizing migration resulted in uneven work loads among the processors. This was due to the fact that the robot performed the task until either the robot achieved the target point or reached an iteration limit. Hence, this migration scheme caused the evolution to wait for the slowest node. In the second stage of the implementation, we attempted to further improve the speedup of the parallel algorithm by the asynchronous migration. When the fastest node reached predetermined generation numbers, the migration request was sent to all subpopulations. Therefore, this scheme caused the evolution of all subpopulations to proceed according to the fastest node. The widely used performance evaluation of the parallel algorithm is the parallel speedup. To make an adequate comparison between the serial algorithm and parallel algorithm, E. Cantú-Paz [2] suggested that the …