GABF: genetic algorithm with base fitness for obtaining generality from partial results: study in autonomous intersection by fuzzy logic

Abstract

Many applications of optimization techniques, such as classification and regression problems, require long simulations to evaluate the performance of their solutions. Problems where the fitness function can be divided into smaller pieces—problem partitioning—demand techniques that approximate the overall fitness from that obtained in a small region of the problem space. This means that less time is spent evaluating individual solutions, which makes such approaches computationally efficient. In this work, a method is proposed to deal with a dynamically calculated fitness function; it is called Genetic Algorithm with Base Fitness (GABF). This method is built over a Genetic Algorithm (GA) to optimize a Fuzzy Rule-Based System (FRBS). The proposed method works by partitioning training data into smaller subsets. The main idea is to assign fitness values derived from part of the training set (or a short simulation) to individuals in the current generation. This fitness value is then inherited and combined with those obtained in subsequent generations. To test the proposal, a scenario in which two vehicles are approaching an intersection is implemented. One vehicle is presumed to be driven by a human and does not change its speed, whereas the other implements an autonomous speed regulator based on fuzzy logic. The regulator must maneuver the autonomous vehicle in a safe and efficient manner. The objective is to optimize both the membership functions and the rule base of the fuzzy system controlling the autonomous vehicle.