A novel Quadtree-Based Genetic Programming for Security and AI Applications

Abstract

Genetic programming (GP) is one of the techniques of the field of genetic and evolutionary computation. Its major task is to evolve computer programs, where each individual is represented as tree structures. Despite its success in multivariate expressions, a different method is needed when the objective has region-by-region properties. In this research, we present a novel Quadtree-based Genetic Programming (Quadtree-GP) for optimizing spatial areas on customized requirements of different grid structures. Instead of using binary trees, the concept of quadtrees are represented as individuals. Three applications of cybersecurity, Machine Learning (ML), and Reinforcement Learning (RL) are introduced for Quadtree-GP: (1) Searchable Encryption in Location-based Alert Systems (2) Pooling in Convolutional Neural Networks, and (3) Reward Shaping in uncertain goal assistance. We denote our methods as Quadtree-GP+, Evolutionary Quadtree Pooling (EQP), and Evolutionary Quadtree-based Reward Shaping (EQRS), respectively. In the first part, we demonstrate that Quadtree-GP+ is able to find searchable encryption candidates that outperform GP search, random search, and Gray Encoding in terms of user response time, token remaining percentage, and execution time. In the second part, we show that the best candidate network of EQP outperforms state-of-the-art max, average, stochastic, median, soft, and mixed pooling in accuracy and overfitting reduction while maintaining low computational costs. In the the third part, we prove that the shaping-reward function generated by EQRS is able to ensure efficient convergence to high-performance policies for the agent and outperforms state-of-the-art Potential-based reward shaping algorithms on uncertain goal environment.