An Evolutionary Algorithm with Variable-Length Chromosome for Multi-objective Minimalistic Attack

Evolutionary Algorithms (EAs) have acquired significant achievements in multi-objective optimization. Canonical EAs are mainly based on fixed-length chromosome. However, certain optimization problems require variable-length chromosome based EAs to solve. In this paper, our interest lies in solving the minimalistic attack problem which is formulated as a multi-objective optimization problem aiming to apply perturbations on the input (pictures) of well-trained deep reinforcement learning (DRL) policies. The objective is to mislead the DRL agent to alter its predictions. To achieve this, we propose a novel evolutionary algorithm with variable-length chromosome that dynamically adapts the chromosome length. Experiments show that the proposed algorithm converges more quickly and achieves better results than the baseline algorithm.