Learning Outcomes
SLO | After this lecture you will be able to... | CLO | Bloom |
|---|---|---|---|
SLO 29.1 | Explain how a genetic algorithm searches with a population, using the terms chromosome, fitness, selection, crossover and mutation | CLO3 | C2 Understand |
SLO 29.2 | Carry out one generation of a genetic algorithm by hand: fitness, roulette selection, one-point crossover and mutation | CLO3 | C3 Apply |
SLO 29.3 | Encode a problem (8-queens, a knapsack load) as chromosomes with a fitness function, and compare a GA with hill climbing and simulated annealing | CLO3 | C3 Apply |
1. Evolution as Search
§1.1 From one climber to a population
Slide L29-S06
Hill climbing (L15) keeps one state and only moves up; simulated annealing (L16) keeps one state but sometimes moves down; local beam search keeps k states. A genetic algorithm (GA) keeps a whole population of answers and makes new answers by combining two parents - so good parts of different answers can be joined. The idea copies natural evolution (Darwin, 1859) and was turned into an algorithm by John Holland (1975). Picture: a farmer who keeps seed from the best cotton plants every season.
Word | Meaning in a GA | Example (x^2 problem) |
|---|---|---|
Chromosome | One candidate answer written as a string | 01101 (x = 13) |
Gene | One position (bit) in the string | the 1st bit |
Fitness | How good the answer is (bigger = better) | f(13) = 169 |
Population | All current answers | 4 strings |
Generation | One round of selection, crossover and mutation | round 1 |
§1.2 The GA loop
Slide L29-S07
population = random strings repeat for G generations: fitness of each string new = [best] # elitism while new is not full: a, b = select(), select() a, b = crossover(a, b) new += [mutate(a), mutate(b)] population = new |
Elitism copies the best answer unchanged into the next generation so it is never lost. A GA is randomised: two runs can give different answers, just like simulated annealing.
2. One Generation by Hand: Maximise f(x) = x^2
§2.1 Fitness and the roulette wheel
Slide L29-S09
x from 0 to 31 is written as a 5-bit binary string (01101 = 8 + 4 + 1 = 13). Fitness = x^2. In roulette-wheel (fitness-proportional) selection each string gets a slice of the wheel equal to its fitness divided by the total fitness (Goldberg, 1989):
No String x f(x)=x^2 Share of wheel Cumulative 1 01101 13 169 0.144 0.144 2 11000 24 576 0.492 0.637 3 01000 8 64 0.055 0.691 4 10011 19 361 0.309 1.000 Total 1170, average 292.5, best 576 |
§2.2 Selection: roulette and tournament
Slide L29-S10
Spin: take a random number r between 0 and 1 and choose the first string whose cumulative share is at least r. Four spins:
Spins [0.1, 0.3, 0.55, 0.8] -> mating pool ['01101', '11000', '11000', '10011'] |
Method | How | Notes |
|---|---|---|
Roulette wheel | Chance = fitness / total | Needs positive fitness; weak when fitnesses are close |
Tournament (k = 3) | Pick 3 at random; the best one is a parent | Simple, robust; used in our 8-queens GA |
Too much selection pressure makes everyone copy the best answer too early, and the variety (diversity) of the population is lost - this is called premature convergence.
§2.3 Crossover and mutation
Slide L29-S11
One-point crossover cuts both parents at the same point and swaps the tails. Mutation flips a gene with a small probability. Real output for the mating pool above:
Cross 0110|1 with 1100|0 at point 4 -> 01100 (12), 11001 (25) Cross 11|000 with 10|011 at point 2 -> 11011 (27), 10000 (16) New generation ['01100', '11001', '11011', '10000'] x = [12, 25, 27, 16] New total 1754, average 438.5, best 729 (was average 292.5, best 576) |
In one generation the average fitness rose from 292.5 to 438.5 and the best from 576 to 729 (x = 27). Mutation is needed because crossover only reshuffles bits that already exist: if no string has a 1 in the first bit, crossover alone can never make one.
Setting | Typical value |
|---|---|
Population size | 20 - 200 |
Crossover probability | 0.6 - 0.9 |
Mutation | small (about one gene per child) |
Elitism | keep the best 1 - 2 |
3. Encoding Real Problems
§3.1 A GA on 8-queens
Slide L29-S13
Chromosome = 8 digits, the row of the queen in each column (as in L15). Fitness = number of non-attacking pairs; there are 8 x 7 / 2 = 28 pairs, so 28 means solved. Mutation moves one queen to a random row.
GA: population 100, tournament of 3, one-point crossover 0.9, mutation 0.3, keep the best 1 GA solved 69 / 100 runs within 300 generations; average 35 generations (about 3509 boards checked) Example solution: 63147025 fitness 28, found in generation 14 |
Method (100 runs) | Solved |
|---|---|
Hill climbing (L15) | 18 |
Genetic algorithm | 69 |
Simulated annealing (L16) | 96 |
The GA is much better than plain hill climbing, but on 8-queens simulated annealing is better: crossover often breaks good boards. No method wins everywhere. GAs shine when an answer is made of parts that can be good on their own.
§3.2 Loading a rickshaw for the mandi (knapsack problem)
Slide L29-S14
A loader rickshaw can carry 500 kg. Chromosome = 10 bits (1 = take the item). Fitness = total profit, or 0 if the load is over 500 kg - a penalty for breaking the rule.
Item kg Rs Wheat bag 100 4200 Cotton bale 170 9500 Rice bag 50 3300 Potato sack 120 3000 Onion sack 80 2800 Mango crate 40 3400 Sugar bag 50 2900 Gur (jaggery) 60 4100 Kinnow crate 30 1500 Fodder bundle 150 2200 Exact (all 1024 loads): 1110011110 500 kg, Rs 28900 GA (population 40, 60 generations), 50 runs: best load found 47 times, average Rs 28806, worst Rs 27300 100 items: 2^100 loads - far too many to check; the GA still checks only about 40 x 60 = 2400 loads. |
The best load (wheat, cotton, rice, mango, sugar, gur, kinnow - 500 kg, Rs 28,900) was found in 47 of 50 runs. With 100 items there would be 2^100 loads - impossible to check - but a GA still checks only a few thousand. Real uses: university timetables, crop plans, delivery routes, and NASA's evolved antenna for the ST5 satellites (2006).
4. Common Mistakes
Wrong idea | Correct idea |
|---|---|
The best string is always selected. | Selection is random; fitter strings only have a higher chance. |
Crossover always makes better children. | Children can be worse; selection removes them later. |
Large mutation explores more, so it is better. | Too much mutation turns the GA into random search. |
A GA always finds the best answer. | It is a heuristic: 69/100 on 8-queens. |
Rule-breaking answers can keep their fitness. | Give them a penalty (here fitness 0). |
Word Bank
Word | Simple meaning | Example |
|---|---|---|
Genetic algorithm | Search that breeds a population of answers | Holland 1975 |
Chromosome | One answer written as a string | 01101 |
Fitness | Score of an answer (bigger = better) | x^2 = 169 |
Selection | Choosing parents, fitter ones more often | roulette, tournament |
Crossover | Swapping the tails of two parents | 11|000 + 10|011 |
Mutation | Small random change to a gene | 10000 -> 10001 |
Elitism | Keeping the best answer unchanged | best 1 |
Penalty | Lower fitness for breaking a rule | overweight load -> 0 |
Premature convergence | Everyone becomes the same too early | lost diversity |
Self-Check Questions
- How is a GA different from hill climbing and beam search? (see §1.1)
- Draw the GA loop and explain elitism. (see §1.2)
- Fitnesses 50, 150, 300: find the roulette shares. (see §2.1)
- How does tournament selection work? (see §2.2)
- Cross 10101 and 01110 at point 2. (see §2.3)
- Why is mutation needed? (see §2.3)
- Give the chromosome and fitness for 8-queens. (see §3.1)
- Why does an overweight load get fitness 0? (see §3.2)
Reading: Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Ch 1-3; Russell & Norvig Ch 4.1.4 (evolutionary algorithms); Lectures 15-16 (local search).