Autonomous selection in evolutionary algorithms

Abstract

This work introduces Autonomous selection in EAs to escape the need for some central control during the selection phases of an EA. The results demonstrate that this is a viable idea that needs further investigation.The main idea is to make the decisions about (de)selection on local level (by the individuals) in a decentralized manner (without global coordination), in such a way that individuals with above/below average fitness have a high/low probability of surviving and producing offspring. The proposed mechanism is based on 1) information about the population's average fitness available at each individual, 2) a function that determines (de)selection probabilities, based on the individual's own fitness and the population's average fitness. This study concentrates on the selection mechanism, and assumes that the average fitness is locally available to all individuals. Indeed, in P2P networks, it is possible to gather an approximation of such statistics without any central control. However, in order to study the selection mechanism in its pure form, it is assumed that an "oracle" provides individuals with the actual average.The parental and survival probabilities of a given individualx are (sigmoid) functions of its fitness deviation from average Δf (x), and depend on two parameters each, sa and ma (with a=s or f, for survival or reproduction). The shift s determines where the transition from low to high probability takes place, and increasing s will increase the selection pressure. The multiplier m determines how sharp the transition from very low to very high probability is. More formally, the probability Pa thatx survives or reproduces isEach individual in turn dies, or survives and reproduces, stochastically with the corresponding probabilities. The population size can hence greatly vary during evolution, as offspring are created locally without any information about global population size.

Evolvable agents, a fine grained approach for distributed evolutionary computing: walking towards the peer-to-peer computing frontiers

“…Within this line, our work (Laredo et al 2007) focuses on self-adaptive population size by the use of the autonomous selection mechanism exposed in Eiben et al (2007) and Wickramasinghe et al (2007).…”

Improving genetic algorithms performance via deterministic population shrinkage

“…Laredo et al expose in [17] the case of a fully distributed EA in which the individuals have to decide on their own state of reproduction without any central control, using instead estimations about the global population state for decision making. The population size varies at run-time as a consequence of such a decentralized reproduction and a self-adjusting mechanism based on autonomous selection [8] tries to keep it stable.…”

Peer-to-peer evolutionary algorithms with adaptive autonomous selection

“…Recent work [7] concerns devising a locally executable function to determine selection probabilities for each individual. Parent selection and survivor selection are separated and handled independently, but selection probabilities in both cases are determined by a sigmoid function.…”

“…Note that by this latter property the population can shrink or grow. This poses a new challenge to the EA designer, because population explosion and implosion should be prevented by calibrating the parameters m and s. Previous work considered a system with perfectly informed individuals that received the exact population statistics from an "oracle" and provided proof-of-principle evidence that the autonomous selection idea is viable [7]. However, tuning m and s required substantial efforts.…”