Identifying similarities in TMBL programs with alignment to quicken their compilation for GPUs
Tony E. Lewis, George D. Magoulas · 2011
The most impressive accelerations of Genetic Programming (GP) using the Graphics Processing Unit (GPU) have been achieved by dynamically compiling new GPU code for each batch of individuals to be evaluated. This approach suffers an overhead in compilation time. We aim to reduce this penalty by pre-processing the individuals to identify and draw out their similarities, hence reducing duplication in compilation work. We use this approach with Tweaking Mutation Behaviour Learning (TMBL), a form focused on long term fitness growth. For individuals of 300 instructions, the technique is found to reduce compilation time 4.817 times whilst only reducing evaluation speed by 3.656%.