Introduction to Genetic Algorithms

• Another Local Search method

• Inspired by natural evolution
Genetic Algorithms Living things evolved into more successful organisms
– offspring exhibit some traits of each parent
Chapter 4.1.4

Introduction to Genetic Algorithms Introduction to Genetic Algorithms

• Keep a population of individuals that are • Mechanisms of evolutionary change:
complete solutions (or partial solutions)
– Crossover (Alteration): the (random)
• Explore solution space by having these combination of 2 parents’ chromosomes during
individuals “interact” and “compete” reproduction resulting in
– interaction produces new individuals
– competition eliminates weak individuals
offspring that have some traits of each parent
• After multiple generations a strong individual
(i.e., solution) should be found • Crossover requires genetic diversity among
• Simulated Evolution via a form of the parents to ensure sufficiently varied
Randomized Beam Search offspring

1
Introduction to Genetic Algorithms Introduction to Genetic Algorithms
• Mechanisms of evolutionary change: • Mechanisms of evolutionary change:
– Mutation: the rare occurrence of errors during – Natural selection: the fittest survive in a
the process of copying chromosomes resulting in competitive environment resulting in better
organisms
• changes that are nonsensical or deadly,
producing organisms that can't survive • individuals with better survival traits
generally survive for a longer period of time
• changes that are beneficial, producing
"stronger" organisms • this provides a better chance for reproducing
• changes that aren't harmful or beneficial, and passing the successful traits on to offspring
producing organisms that aren't improved • over many generations the species improves
since better traits will out-number weaker ones

Genetic Algorithm
Representation of Individuals
Create initial random population
Solutions represented as a vector of values
– For example: Satisfiability problem (SAT) Evaluate fitness of each individual
• determine if a statement in propositional logic is yes
satisfiable, for example: Termination criterion satisfied?
(P1∨ P2)∧(P1 ∨ ¬P3)∧(P1 ∨ ¬P4)∧(¬P3 ∨ ¬P4) stop
• each element corresponds to a symbol having no
a value of either true (i.e., 1) or false (i.e., 0) Select parents according to fitness
• vector: P1 P2 P3 P4
• values: 1 0 1 1 ç rep. of 1 individual Combine parents to generate offspring
– Traveling salesperson problem (TSP)
• Tour can be represented as a sequence of cities visited Mutate offspring

Replace population by new offspring

2
 Genetic Algorithm (1 version*)
Initialization: Seeding the Population
1. Let s = {s1, …, sN} be the current population
2. Let p[i] = f(si)/SUMjf(sj) be the fitness probabilities • Initialization sets the beginning population
3. for k = 1; k < N; k += 2 of individuals from which future generations
• Parent1 = randomly pick si with prob. p[i] are produced
• Parent2 = randomly pick another sj with prob. p[j]
• Randomly select 1 crossover point, and swap • Issues:
strings of parents 1 and 2 to generate two children – size of the initial population
t[k] and t[k+1] • experimentally determined for problem
4. for k = 1; k ≤ N; k++ – diversity of the initial population (genetic
• Randomly mutate each position in t[k] with a diversity)
small probability • a problem resulting from lack of diversity is premature
5. New generation replaces old generation: s = t convergence to a non-optimal solution
*different than in book

Initialization: Seeding the Population Evaluation: Ranking by Fitness

• How is a diverse initial population generated? • Evaluation ranks the individuals using some
– uniformly random: generate individuals fitness measure that corresponds with the
randomly from a solution space with uniform distribution
quality of the individual solutions
– grid initialization: choose individuals
at regular "intervals" from the solution space
• For example, given individual i:
– non-clustering: require individuals to be a predefined
"distance" away from those already in the population – classification: (#correct(i))2
– local optimization: use another technique (e.g. HC) – TSP: 1/tour-length(i)
to find initial population of local optima; doesn't ensure – SAT: #ofClausesSatisfied(i)
diversity but guarantees solution to be no worse than the
local optima

3
 Selection: Finding the Fittest Selection Techniques
• Choose which individuals survive and possibly • Deterministic Selection
reproduce in the next generation – relies on evaluation/fitness function
• Selection depends on the evaluation/fitness – converges fast
function • Two approaches:
– if too dependent, then, like greedy search, a non- – next generation contains parents and children
optimal solution may be found • parents are the best of the current generation
– if not dependent enough, then may not converge to • parents produce children, and parents survive to next
a solution at all generation
• Nature doesn't eliminate all "unfit" genes; – next generation contains only children
they usually become recessive for a long period • parents are the best of the current generation
of time, and then may mutate to something • parents are used to produce children
• parents don't survive (counters early convergence)
useful

Selection Techniques Selection Techniques

• Proportional Fitness Selection
– each individual is selected proportionally to their
fitness score Proportional selection example:
– even the worst individual has a chance to survive • Given the following fitness values for population:
– helps prevent “stagnation” in the population l Sum all the Fitnesses
Individual Fitness Prob.
• Two approaches: 5 + 20 + 11 + 8 + 6 = 50
A 5 10%
– rank selection: individual selected with a l Determine probability B 20 40%
probability proportional to its rank in population for each individual, i
sorted by fitness C 11 22%
Fitness(i) / 50 D 8 16%
– proportional selection: individual selected with a
probability: Fitness(individual) / ∑ Fitness for E 6 12%
all individuals

4
 Selection Techniques Crossover: Producing New Individuals
• Crossover is used to produce new
Crowding: a potential problem associated with individuals (i.e., children)
Selection • Crossover for vector representations:
– occurs when the individuals that are most-fit – Pick pairs of individuals as parents and randomly
quickly reproduce so that a large percentage swap their segments
of the entire population looks very similar – also known as "cut and splice"
– reduces diversity in the population
– may hinder the long-run progress of the algorithm • Parameters:
– number of crossover points
– positions of the crossover points

Crossover: Producing New Individuals Crossover: Producing New Individuals

• 1-point Crossover • N-point Crossover

– generalization of 1-point crossover
– pick a dividing point in the parents' vectors
– pick n dividing points in the parents' vectors and
and swap their segments splice together alternating segments
• Example • Uniform Crossover
– given parents: 1101101101 and 0001001000 – the value of each element of the vector is
randomly chosen from the values in the
– crossover point: after the 4th digit corresponding elements of the two parents
– children produced are:
• Techniques also exist for permutation
1101 + 001000 and 0001 + 101101 representations

5
 Genetic Algorithm (1 version*)
Producing New Individuals
1. Let s = {s1, …, sN} be the current population
2. Let p[i] = f(si)/SUMjf(sj) be the fitness probabilities
• Mutation 3. for k = 1; k < N; k += 2
– randomly change an individual • Parent1 = randomly pick si with prob. p[i]
• Parent2 = randomly pick another sj with prob. p[j]
– e.g. TSP: two-swap, two-interchange
• Randomly select 1 crossover point, and swap
– e.g. SAT: bit flip
strings of parents 1 and 2 to generate children t[k]
and t[k+1]
• Parameters: 4. for k = 1; k ≤ N; k++
– mutation rate • Randomly mutate each position in t[k] with a
– size of the mutation small prob.
5. New generation replaces old generation: s = t
*different than in book

GA Solving TSP
Genetic Algorithm Applications

6
 Genetic Algorithms as Search Genetic Algorithms as Search
The Problem of Local Maxima
Individuals get stuck at pretty good, but not • GA is a kind of hill-climbing search
optimal, solutions • Very similar to a randomized beam search
– any small mutation gives worse fitness • One significant difference between GAs and HC
– crossover can help get out of a local maximum is that, it is generally a good idea in GAs to fill
– mutation is a random process, so it is possible that the local maxima up with individuals
we may have a sudden large mutation to get these • Overall, GAs have less problems with local
individuals out of this situation maxima than HC or neural networks

Summary
• Easy to apply to a wide range of problems
– Optimization problems such as TSP
– inductive concept learning
– scheduling
– Layout
• Results can be very good on some problems and
poor on others
• GA is very slow if only mutation is used;
crossover makes the algorithm significantly
faster