Genetic Programming Multitasking.

In this paper, we present a new multitasking algorithm for Genetic Programming (GP). Our proposed algorithm (referred to as "GP-Tasking") evolves population using multifaceted strategy. Each individual is trained with different training sets and evaluated with multiple fitness functions (where each fitness function represents one task). At the beginning of the run, GP-Tasking, randomly uses crossover operator to facilitate knowledge transfer between different tasks and store probability of constructive crossover operators between different tasks. This information is used to bias the crossover between tasks that have higher probability of producing fitter offspring. The novelty of GP Tasking, is that it uses one population in the same phenotype space but with different interpretations to explore multiple genotype spaces. GP-Tasking was evaluated with 3 sets of experiments where in each set we tested GP-Tasking ability to solve 5 different tasks simultaneously. Results showed that GP-Tasking evolved smaller solutions and consume significantly less computational time.