CodinGame — decisions under constraints
Competitive game AI explored through C++ bots, opponent modelling and controlled experimentation.
The engineering problem
A game bot has to choose a useful action from an evolving game state within a limited computation budget. That turns algorithm design into a practical trade-off between search depth, evaluation quality and execution time.
Model the opponent, then check the assumption
The Fall 2026 work explored mixtures of opponent forecasts. Variants were compared on recorded positions and against the bot’s own previous versions. Results on self-play and real opponents did not always agree, which made the evaluation method as important as the heuristic itself.
Make the measurement trustworthy
A review of the evaluation bench found gaps in reconstructed positions and opponent transitions. Correcting the bench and checking that a version compared against itself produced no difference came before interpreting new improvements.
That is the central lesson of this project: an apparent gain only matters if the test actually represents the decision being made.
Work within the time budget
Iterations balanced prediction quality with a safety margin for turn time. More complex forecasts were not automatically better, and some variants were rejected when the extra complexity did not produce convincing gains.
Results and scope
The Fall notes record Legend league and a ranking of 145 out of 1,047. That is a project-note record, not a live ranking verified here. Intermediate arena ranks are intentionally not presented as final results.
The implementation and experiment logs are private. The transferable work is the process: establish a baseline, verify the bench, compare candidate decisions and keep the limitations visible.