Fast, comfortable or economical: Evolving platooning strategies with many objectives

Abstract

The research in this paper is inspired by a vision of intelligent vehicles that autonomously move along motorways: they join and leave trains of vehicles (platoons), overtake other vehicles, etc. We propose a multi-objective evolutionary algorithm that evolves high-level controllers for such intelligent vehicles. The algorithm yields a set of solutions that each embody their own prioritisation of various user requirements such as speed, comfort or fuel economy. This contrasts with the current practice in researching such controllers, where user preferences are summarised in a single number that the controller development process then optimises.In this article, we test our multi-objective approach on 6 objectives. Our method outperforms a widely used human behavioural model on many of the objectives. Some performance is lost when we introduce additional objectives, but these losses are small and therefore acceptable. We show that it is possible to evolve a set of vehicle controllers that correspond with different prioritisations of user preferences, giving the driver, on the road, the power to decide which preferences to emphasise, although we do see that the more objectives are added to the system, the less intuitive the prioritisation of the different objectives becomes.

Coping with opponents: multi-objective evolutionary neural networks for fighting games

“…So far, only a few multi-objective approaches can be found in the literature: The modular multi-objective NEAT (MM-NEAT) [8] which evolves modular ANNs based on NSGA-II [9] and NEAT. Van Willigen et al proposed the NEAT-PS [10] that utilizes the Pareto strength approach of SPEA2 [11] to aggregate a vector of fitness values into a single scalar and make it applicable in standard NEAT. Furthermore, NEAT-MODS [12] is a multi-objective version of NEAT that promotes an elitist and diverse population by a modified selection procedure.…”

“…MM-NEAT is based on NSGA-II with components borrowed from NEAT [8]. NEAT-PS Van Willigen et al [10] proposed the NEAT-Pareto strength (NEAT-PS) which falls back to another well-known EMOA, the SPEA2 [11]. Here, the authors mainly utilize the Pareto strength approach which aggregates a fitness vector into a scalar fitness value for each solution in a given set.…”

“…Using that scalar, any multi-objective optimization problem can be treated as a single-objective one in NEAT-PS. Thus, the only addition to NEAT is the Pareto strength fitness determination in NEAT-PS [10]. NEAT-MODS Based on investigations of NEAT-PS that revealed that the algorithm does not guarantee a uniform evolution of all objectives due to its focus on elitism only, Abramovich and Moshaiov proposed the NEATmulti-objective diversified species (NEAT-MODS) [12].…”

“…In particular, NEAT-PS seems to depend strongly on an adequately large population. One major problem occurring with NEAT-PS might be its difference to SPEA2: While SPEA2 relies on Pareto strength in combination with a diversity measure [11], NEAT-PS does only apply the Pareto strength approach without respect to diversity [10]. Future research will also investigate approaches for dynamically adjusting the population size.…”

A smart integrated environment for vehicular traffic simulation

“…In another study, a marathon competition has been simulated with the help of MovSim . MovSim has been used for simulation of autonomous vehicle agents enabled with a multiobjective method to find multiple prioritizations for passengers traveling in an autonomous vehicle so that the passenger may chose the way car should be drove . To incorporate human factors in microscopic driver models, a model of personality profile has been incorporated in existing microscopic models, and simulation‐based experiments have been performed using MovSim …”