Research / FPL optimization
Evolving an FPL squad
Selection, crossover, mutation, and the search for a legal team
An FPL squad can be represented as a small string of choices. Genetic algorithms treat many such strings as a population. The strongest squads survive, exchange parts, mutate, and produce a new generation.
This thesis tests ten versions of that process on historical 2023/24 data. Its most useful result is not one perfect team. It is a demonstration of how the search rules change what the algorithm can discover.
Start with a population
The algorithm creates many candidate squads, then scores each with a fitness function based on player performance, prices, positions, and FPL rules. Better squads receive a greater chance to influence the next generation.
The process does not examine every legal team. It searches by preserving and recombining parts of teams that already look useful.
Selection, crossover, mutation
Three operations shape the search. Selection decides which squads reproduce. Crossover joins parts of two squads. Mutation changes individual choices to prevent the population from becoming identical too early.
The thesis compares rank, roulette, and tournament selection; one-point, two-point, and uniform crossover; and scramble, swap, and inversion mutation.
- 01Score every legal candidate
- 02Select stronger parents
- 03Recombine player choices
- 04Mutate and repeat
A search method needs enough discipline to improve and enough disorder to escape its first answer.
nil nil interpretation
The winning recipe
The best reported configuration used tournament selection, uniform crossover, and scramble mutation. Tournament selection repeatedly compares small groups and advances the strongest. Uniform crossover chooses each component from either parent. Scramble mutation rearranges a selected subset.
Within this experiment, that combination reached the highest fitness most consistently. The result describes the search setup, not a universal FPL strategy.
The fitness function defines success
A genetic algorithm only becomes as intelligent as its fitness measure. If fitness rewards past points per game, the search finds teams that would have looked strong by that measure. It does not automatically understand future minutes, transfers, captaincy, or fixture runs.
The right research question is therefore not whether evolution can find a high-scoring squad. It is whether the score guiding evolution predicts future FPL value without using information from the future.
Optimization, not proof of prediction
This 2024 master's thesis evaluates genetic configurations using historical 2023/24 data. The reported fitness gains show that one search setup optimized its objective better than the others. They do not establish that the resulting squads would win prospectively, especially if the fitness inputs include season information unavailable at the decision time.
Read the original research ↗