Checklist EUROAI (CEOAI) 2026 Practice Round 1 · Task 1
The Stochastic Rift
Estimate optimal state values of an unknown stochastic MDP from a fixed log of about 6,000 shuffled transitions (offline reinforcement learning).
The task
Sector 7 is a stochastic MDP with discrete states 0…N−1, 4 actions, unknown noisy transitions and rewards, and discount γ = 0.99. Only a static 'black box' flight log of a failed drone is available (current_state, action, reward, next_state), non-sequential and sparse; the same state–action pair can lead to very different outcomes.
For each query state the contestant must estimate the true optimal value V*(s), which is computed by a ground-truth solver with access to the exact transition matrix. An env.py file defines the state and action spaces and a step() method for local simulation.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
sector_logs.csv(~6,000 transitions),query_states.csv(id,state_id),env.py.- You submit
predictions.csvwith subtaskID (1), datapointID (query id) and answer (estimated V*).- Scoring
- MSE against V*. Points = 100 × ((4000.0 − MSE) / (4000.0 − 205.0))^4, 0 if MSE ≥ 4000, capped at 100 if MSE ≤ 205.
- Rules
- Solution within 30 minutes on a standard single-core CPU.
- External data strictly prohibited.
- Format
- Online Practice Round 1, 15 March 2026, 13:00–19:00 UTC (6 hours), open to everyone on Nitro AI Judge; tasks proposed by volunteers and reviewed by the International Scientific Committee.