Download Evolutionary Computation Introduction

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Sociobiology wikipedia , lookup

Genome (book) wikipedia , lookup

Adaptive evolution in the human genome wikipedia , lookup

Transcript
Evolutionary
Computation
Introduction
Peter Andras
[email protected]
www.staff.ncl.ac.uk/peter.andras/lecture
s
Overview
1. Biological inspiration
2. Artificial genes
3. Learning by evolution
4. Artificial evolution
5. Learning by artificial evolution
Biological inspiration
Evolution:
• Darwin
• from bacteria to sponges, insects, fishes, and mammals
• from simple organs to complex ones
• from randomly spread neurons to highly organized large
brains
Biological inspiration
Foundations:
• nucleic acids: adenin, citozin, guanin, timin, uracil
• DNA
• chromosomes
• genes
• RNA, proteins, cells
Biological inspiration
Adaptation by evolution:
• ecological niche: a set of ecological conditions
(e.g., food resources, predators, other
environmental risks, threats and opportunities);
• conquering new ecological niches (e.g., islands)
• development of new species that are able to use
the opportunities provided by a new niche and
avoid the related dangers;
Biological inspiration
Adaptation:
• development of new behaviours and organs;
• new cells and cell behaviours;
• new proteins;
• new genes;
Artificial genes
Idea:
• copying natural evolution by emulating genes and
their evolution;
Objective:
• developing adaptive solutions of some problems;
Artificial genes
Artificial world:
• world of problems;
Artificial individuals:
• solutions of the problems
• genes encode features of the problem solutions
Artificial genes
Discrete feature encoding:
• e.g., 0 and 1 for the presence or absence of the
features;
• chromosomes: 001110101110;
• the genes do not represent necessarily full features;
Artificial genes
Continuous type feature encoding:
• e.g., features encoded by real numbers;
• chromosomes: multi-dimensional real vectors;
• usually genes directly encode features;
Learning by evolution
Learning:
• learning = adaptation
• adaptation = optimisation
• optimisation criteria: fitness in the given
environmental conditions;
Learning by evolution
Exchanging and combining genes:
• sexual crossover
+

Learning by evolution
Mutation:
• random changes of the genes

Learning by evolution
Inheritance:
• the offspring inherits the properties of their parents;
• some combinations are lethal;
• the inherited properties range from similar proteins to
similar behaviours;
Learning by evolution
New species:
• slow evolution;
• accumulating minor changes;
• modifications of organ functionality;
• selection of some variants of standard features (e.g.,
feather colours);
• emergence of new behaviours, organs;
Learning by evolution
Mating success:
• features that better fit the environmental niche increase
the chance of the individual to get mates and reproduce;
• individuals with higher fitness have more offspring;
• the genes of the successful individuals spread within
the population and become dominant;
• genes that cause evolutionary advantage in mutated
individuals become general;
Learning by evolution
Evolutionary optimisation:
• increased fitness in the ecological niche;
• mutation is responsible for new genes (proteins, cells,
organs, behaviours);
• crossover is responsible for passing over the new
genes;
• fitness based mating success is responsible for the
emergence of domination of genes that increase fitness;
Artificial evolution
Evolution of a population of problem solutions:
• individuals are the problem solution;
• each solution is characterized by its features encoded
by the genes;
• evolution by genetic operators and offspring
generation;
Artificial evolution
Mutation operator:
• randomly change the genes encoding the solution
features;
• e.g., changing a 0 into a 1 and inversely;
• e.g., minor modification of a feature encoded by a real
number;
Artificial evolution
Crossover operator:
• defines how to select exchanged parts of the
genetic material;
• e.g., randomly selecting a chromosome splitting
position;
Artificial evolution
Directed operators:
• preferential selection of some genes for mutation or
some segments of the chromosome for crossover;
• the preferential selection is based on monitoring,
which components of the solution contribute to bad or
good performance;
Artificial evolution
Constrained operators:
• mutation constraints: some simultaneous mutations are
not allowed, others are enforced;
• crossover constraints: some chromosome segments are
allowed to be exchanged only for some chromosome
segments with specified location;
Artificial evolution
Optimisation energy function:
• fitness measure = problem solving performance
• problem solving performance of the individuals are
evaluated with a random sample of the potential
problems;
Artificial evolution
Mating potential:
• it is based on the problem solving performance;
• the number of the offspring of the individuals
depends on their mating potential;
• high fitness individuals have many offspring that
inherit at partly their features;
Artificial evolution
Many parent mating:
• the crossover applies to the mix of all parents;
Learning by artificial
evolution
Problem solving performance optimisation:
• the average performance of the population increases;
• the best performing individuals represent very good
solutions after long enough evolution;
Learning by artificial
evolution
Key features:
• proper feature coding;
• proper evolutionary operators;
• proper fitness evaluation;
• proper mating selection;
Learning by artificial
evolution
Feature coding:
• the important solution features should be encoded;
• if it is not clear what is important and what is not,
better to encode more features than less features;
• the feature coding and the decoding of the code
should not be ambiguous;
Learning by artificial
evolution
Evolutionary operators:
• the result of applying evolutionary operators should be
meaningful;
• the crossover should result individuals that inherit their
parents properties;
Learning by artificial
evolution
Fitness evaluation:
• the fitness function should be closely related to the
effective problem solving performance;
Learning by artificial
evolution
Mating potential determination:
• the more fit individuals should have more offspring;
• the drastic elimination of less fit individuals may lead
to the elimination of genes that are sleeping but may
become important for the achievement of very high
performance;
Learning by artificial
evolution
Problems:
• too narrow spread of performances: it is likely that
there is little genetic variation in the population;
• too large spread of the performances: it is possible that
the encoding of features or the genetic operators are not
functioning properly;
• too slow increase of the average performance: it is
possible that the encoding of features or the genetic
operators are not functioning properly;
Summary
• evolution leads to niche adapted new species;
• the basis of evolution are the genes;
• new genes may lead to new proteins, cells, organs,
behaviours, which may increase the fitness of the biological
organism;
• evolutionary adaptations spread by mating and by higher
mating success of those who are more fit to the ecological
niche;
• evolutionary learning means optimisation of the fitness;
Summary
• artificial genes encode features of solutions of some
problems, the encoding can be discrete or continuous;
• artificial evolution works by genetic operators;
• genetic operators: mutation, crossover, directed operators,
constrained operators;
• mating potential depends on problem solving performance;
• having appropriate feature encoding, evolutionary
operators, fitness function and mating potential
determination, the artificial evolution leads to high
performance solutions of the problem;