Research / FPL strategy
Can an AI beat 2.5 million managers?
An early automated FPL manager reached around the top percentile
Before FPL models became an industry, three researchers built an automated manager and entered it against 2.5 million people. It did not know injury news, press conferences, or the private information available to human managers. It still finished around the top percentile.
The result came from treating FPL as a sequence of connected choices. A transfer was not judged only by the next score. It changed the squad, budget, and options available in every later gameweek.
A season, not a lineup
Choosing the best available eleven for one week is already a constrained problem. The manager must respect budget, position, club, and formation rules. Across a season, the search becomes much larger because each transfer changes the next decision.
The paper counted more than 500 selectable footballers and more than 1025 legal squad combinations. The automated manager had to make this choice thirty-eight times while paying four points for extra transfers.
The best gameweek team can be the wrong season-long decision.
00.football interpretation
Beliefs instead of certainty
The system represented each player's ability as a probability distribution. New match results updated those beliefs. A strong score could raise the estimate, but the model retained uncertainty rather than treating the latest return as a permanent truth.
This matters because uncertainty has value. A manager can favor a reliable player when protecting against a bad outcome, or explore a less certain option when the possible gain justifies it. The paper used Bayesian Q-learning to connect those beliefs with future rewards.
- 01Estimate player and team ability
- 02Sample possible match outcomes
- 03Construct legal squads
- 04Value the present and future return
The machine held its own
On data from the 2010/11 season, the Bayesian approach produced the highest mean final score among the tested methods. Its average result sat around the top percentile. Its best simulated season reached 2,222 points, which would have placed inside the top 500.
The important comparison was not machine against a perfectly informed human. The agent lacked information that real managers could use. Its advantage came from disciplined treatment of uncertainty and the transfer horizon.
The model did not predict football perfectly. It avoided making every decision as if the next week were the last.
00.football interpretation
What still holds
The rules, prices, player pool, and information environment have changed since 2010/11. The computing methods are old by current standards. The strategic lesson remains current: FPL rewards a policy, not a list of picks.
A useful decision model must price future transfers, uncertainty, and flexibility. A high expected score next week is only one part of the value of owning a player today.
One old season, one early benchmark
This peer-reviewed AAAI paper evaluated its agents on the 2010/11 FPL season. The comparison was simulated from historical results rather than a live prospective entry, and the model simplified some parts of football performance. Its rank claims should be read as a benchmark for the tested season, not proof that the same method would dominate modern FPL.
Read the original research ↗