Fighting for #1 in the Ultimate Tic-Tac-Toe Arena

You can now submit highly optimized AI models to competitive programming platforms, bypassing language limitations.

CodinGame's Ultimate Tic-Tac-Toe (UTTT) bot programming arena, launched in March 2018, has become highly competitive, with the top rank fiercely contested. The arena imposes constraints on submissions, including a 100k character limit, a single vCPU, and a 100 ms turn limit, which unexpectedly level the playing field by preventing massive, resource-intensive AI models. UTTT's ruleset is elegant and not yet solved, making it an ideal candidate for Monte Carlo Tree Search (MCTS). However, top-tier performance requires more advanced techniques like neural network evaluation functions (NNUE), which are significantly more effective than traditional heuristics due to UTTT's complex positional evaluations. The article details the implementation of NNUEs, including optimizations like side-to-move encoding and output bucketing, and explains how a Python submission can deliver a compressed, compiled C binary to bypass Python's performance limitations and meet the character count requirement.

The strict constraints of the CodinGame UTTT arena, initially appearing limiting, have proven to be an ingenious feature. By capping code size and compute per turn, the platform democratizes high-level AI competition, enabling individual developers with standard hardware to compete against sophisticated algorithms. This design choice fosters innovation by requiring clever optimization and efficient algorithms rather than brute-force computation.

The success of NNUEs in UTTT highlights a broader trend in AI development where efficiently updatable neural networks are becoming crucial for complex strategy games. The ability to achieve top performance within tight resource constraints, as demonstrated by the compressed C binary delivered via Python, points to a future where performance engineering and clever deployment are as important as the AI model itself.