FPL

Description

FPL is an ORS environment that simulates the Fantasy Premier League game. Agents manage a fantasy football team across a full 38-gameweek Premier League season (2024-25 or 2025-26), developing and executing ML-based strategies for squad selection, transfers, captaincy, and formation decisions.

Capabilities

• Developing machine learning models for player performance prediction
• Squad selection and management within a £100M budget
• Transfer strategy and timing optimization
• Captain selection and starting XI formation decisions
• Strategic chip usage (Wildcard, Bench Boost, Triple Captain)
• Long-horizon multi-turn execution (38 gameweeks)

Compute Requirements

Agents in FPL are given a sandbox with 4 CPUs and 4GB RAM, outbound network access blocked, and a Python 3.12 data science image.

Tasks

There are two training tasks in this environment:

• 2024-25 Season: The agent manages a fantasy team for the full 2024-25 Premier League season (38 gameweeks).
• 2025-26 Season: The agent manages a fantasy team for the full 2025-26 Premier League season (38 gameweeks).

Each gameweek, the agent must:

1. Select a 15-player squad (2 GK, 5 DEF, 5 MID, 3 FWD) within a £100M budget, with a maximum of 3 players per club.
2. Choose a captain (2x points) and vice-captain (fallback if captain doesn't play).
3. Pick a starting XI with a valid formation (1 GK, 3-5 DEF, 2-5 MID, 1-3 FWD) and set 4 substitutes in priority order.
4. Optionally make transfers (1 free per gameweek, accumulating up to 5; additional transfers cost 4 points each).
5. Optionally play a chip (one per season: Wildcard, Bench Boost, or Triple Captain). Chip rules are identical across seasons as a deliberate simplification — the real 2025-26 game granted a second set of chips at gameweek 20, which this environment does not model.
6. Advance to the next gameweek to receive results.

Reward Structure

This is a dense, verifiable reward environment. The primary reward is returned after each gameweek via the next_gameweek tool. Smaller rewards are also given for completing squad selection (make_initial_transfers returns reward 1.0) and transfer penalties are reflected immediately (make_transfer returns the negative penalty as reward). The gameweek reward is:

$\text{GW Points} = \sum_{i \in \text{Final XI}} \text{score}_i + \text{Bench Boost}$

Where each player's score is their base gameweek points, except the captain whose score is multiplied by 2 (or 3 with the Triple Captain chip).

• Bench Boost: If active, bench players' scores are added to the total.
• Transfer Penalty: 4 points per transfer beyond free transfers, charged as a negative reward by make_transfer at the time of the transfer rather than deducted from the gameweek score (waived with Wildcard chip).
• Automatic Substitutions: If a starting player doesn't play (0 minutes), the highest-priority eligible substitute replaces them.

If the captain doesn't play, the vice-captain automatically receives the captain multiplier instead.

We do not use LLM graders for this task.

Data

Agents are given access to historical player data for the task season, including:

• Player statistics (names, positions, clubs, values, total points)
• Fixture schedules with kickoff times
• Team information
• Individual player gameweek data (points, minutes played, goals, assists, etc.)

Agents also have a bundle of prior-season data (2016-17 up to the season before the task season) mounted read-only at /tmp/gr-datasets for building ML strategies. In addition, after each gameweek advance the environment delivers a fresh copy of the current season's data so far — per-player gameweek stats and finished fixtures up to (but never including) the upcoming gameweek, plus each player's prior-season history — extracted at the fixed path /home/ubuntu/latest_data/latest/ (and left as a zip alongside it at /home/ubuntu/latest_data/latest.zip), so agents can retrain on results as they unfold. The path does not change between gameweeks: the tree is replaced in place on every advance, so a pipeline that re-reads it each gameweek always sees the freshest data. The sandbox has outbound network access blocked so the agent cannot look up the task season's real outcomes; the weekly data is fetched server-side and pushed into the sandbox. Data is sourced from the vaastav/Fantasy-Premier-League dataset.

Tools

Agents are given access to CLI tools for creating, viewing, and searching a filesystem. They are also given environment-specific tools:

• list_players: Browse players by position with pagination
• make_initial_transfers: Select the initial 15-player squad
• make_transfer: Transfer players in/out after gameweek 1
• set_captain: Choose captain and vice-captain
• pick_starting_xi: Set starting XI and substitute priority order
• view_current_squad: View current squad status, captains, and formation
• play_chip: Activate a chip (Wildcard, Bench Boost, or Triple Captain)
• list_fixtures: View fixtures for a given gameweek
• next_gameweek: Advance to the next gameweek and receive results

Time Horizon

FPL is an open-ended, long-horizon environment where agents play through an entire 38-gameweek Premier League season. Each gameweek requires multiple tool calls for squad management (transfers, captaincy, formation) before advancing.

[Statistics on average tool calls here]

Environment Difficulty

[Statistics on environment difficulty here]

Safety

Agents in FPL are told to maximize their total season points. The environment does not present direct safety risks, as agents only interact with a simulated Fantasy Premier League game with no real-world financial transactions. All decisions are made against historical data within a sandboxed environment.

We think it is unlikely that optimising against the objective of maximising points in this game would promote unaligned behaviour through RL, as there is no opportunity to manipulate or hurt others in the course of maximising this objective in the game.

Citations

@dataset{GRFPL,
  author    = {General Reasoning Inc. Team},
  title     = {FPL},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/FPL}
}