Detailed information for this page is currently under preparation. The content will be announced and updated on this website once it is officially confirmed.
Please check this page regularly for future updates.
AI in Games Contest 2026 invites participants to design intelligent agents capable of strategic decision-making, adaptation and autonomous competition in the Commander game environment.
The official competition date and schedule will be announced soon.
Detailed venue information will be published once confirmed.
Registration procedures and submission instructions will be announced soon.
The contest is based on Commander, a real-time strategy environment with turn-based tactical elements. The game is played on a hexagonal map in which players expand territory, occupy strategic locations, accumulate forces and attempt to locate and capture their opponents' headquarters.
Participants do not directly control the game. Instead, each team develops an autonomous intelligent agent that observes information available within its own field of vision and makes decisions during each game step.
Successful agents must combine strategic expansion, resource management, navigation, defence and decision-making under incomplete information.
Decide when and where to expand, defend territory, move forces and engage opponents.
Agents can only use information visible from their own controlled areas, requiring decisions under uncertainty.
Once a match begins, intelligent agents independently interact with the game environment through the provided interfaces.
Participants may develop rule-based agents or use machine-learning and reinforcement-learning approaches.
The contest platform provides the tools required for agent development, testing, evaluation and AI training.
| Game Environment | Strategic competition on a hexagonal-grid map. |
|---|---|
| Player Objective | Expand territory, accumulate forces and capture opponents' headquarters. |
| Information | Agents operate under limited visibility and may only use information available within their own field of vision. |
| Agent Languages | Lua and Python. |
| Development | Rule-based strategies, search algorithms, machine learning and reinforcement learning approaches may be developed using the provided interfaces. |
| Evaluation | Agents can be tested through individual matches and repeated automated evaluations before the competition. |
The platform includes example agents demonstrating basic exploration, greedy expansion and path-finding strategies.
Participants can build upon these examples and introduce defence, force allocation, target prioritisation, historical state, opponent modelling and AI-based decision systems.
The training environment provides observations, legal-action masks, rewards and episode termination information through a dedicated Python training interface.
Example training workflows include DQN, MaskablePPO and two-agent self-play.
Contest software, technical documentation and additional materials will be published here as they become available.
The current Commander II desktop package is designed for Windows x64. Participants should fully extract the package to a writable local directory before running the application.
Windows systems may require .NET Framework 4.6.2 or later and Microsoft Edge WebView2 Runtime. These components are normally available on current Windows 10 and Windows 11 systems.
The competition date, venue and registration procedure are currently being finalised. Updated information will be published on this page once confirmed. Participants are encouraged to check this page regularly for the latest announcements and additional contest materials.