1: \begin{abstract}

2: Uniform Addition by Mutation and Deletion (UMAD) has been shown to outperform other variation operators spanning both mutation and crossover on software synthesis problems. In its standard form, UMAD adds new genes into a genome at a predetermined rate, and then deletes genes from the augmented genome at a rate that ensures the expected size of the genome does not change. While this predetermined mutation rate already outperforms many other mutation and crossover schemes, we do not expect a preset mutation rate to be optimal across all software synthesis problems and across all generations within one run of a genetic algorithm. However, many current adaptive mutation rate schemes such as self-adaptive mutation rates suffer from the vanishing mutation rate problem, in which the mutation rate quickly decays to 0, causing premature convergence. To address this, we propose an adaptive scheme based on max-bandits and tile-codings to find the optimal mutation rate. Results on software synthesis problems validate the effectiveness of our approach, as we achieve significantly better results than vanilla UMAD.  

3: 

4: 

5: 

6: 

7: 

8: \keywords{Uniform Addition by Mutation and Deletion  \and Adaptive Mutation Rate \and Genetic Programming \and Software Synthesis}

9: \end{abstract}

10: