How do I start training an AI commander?
BotBellum is an AI training game where your commander makes the decisions during automated ranked fights. Your role is to build it, run experiments and decide which version to publish.
The way commanders are trained is being reworked, so the available training controls may change. This guide focuses on the core game loop and reading results. It does not depend on a particular training algorithm.
Build a commander
Sign in with a free account, open My Commanders and build a commander. Give it a name and choose its brain, the structure it uses to make decisions.
Run training
Open Training for that commander. Review the options shown in the workspace and start a run against available compatible published commanders.
Review the evidence
Open Analytics and choose Training. Inspect a completed generation and its recorded fights before deciding what to try next.
Publish a version
Publish the commander when you want it to enter ranked competition. Training again does not replace that published version until you publish the changes.
What should I focus on in my first experiment?
Start with one observable question. Is the commander leaving a lane undefended, spending energy without taking structures, or winning against only a narrow set of opponents? Available replays can help you choose a question to investigate.
Doctrine describes scoring preferences such as pressure, defense and efficient trades. The current training implementation determines how those preferences are applied. A higher score for one preference does not by itself show that a commander wins more fights.
As an illustrative experiment, suppose a replay shows the commander attacking while its own core is threatened. Record that behavior, change one relevant option available in Training, then inspect whether the next recorded fights show the same problem. This is an example of a test to run, not a measured result or a guaranteed fix.
Does AI training run in my browser?
Yes. Training currently runs on your device while BotBellum is open. Keep the game open during a run. Training speed depends on your device and the work being evaluated.
A run contains generations, with candidate versions evaluated through fights. Training needs compatible published opponents before it can begin. If none are available, the workspace cannot start that run.
You can also fork a published commander when its creator allows it. The fork becomes your own commander in My Commanders, where you can train it further. Its source commander’s ranked record does not become yours.
How can I tell whether training helped?
In Analytics, choose Training and select a run and generation. Results, Units and lanes, Timeline, and Fights describe the selected evidence. Follow the latest completed generation or keep one selected while training continues.
Check fight counts, opponents and game rules before comparing outcomes. Read the recorded candidate identity too. The saved version can include an update after the measured fights, and the retained checkpoint’s final evaluation is shown separately.
Do not compare raw training fitness across different scoring objectives or training implementations as though it were a common scale. Use compatible recorded fight outcomes and behavior. A sample replay explains one fight, not every fight in a generation.
Why can training and ranked results differ?
Training results describe the candidate versions and opponents in that run. Ranked competition uses your published version against the opponents it encounters there. Different versions, opponents and sample sizes can produce different results.
Publish the version you want to compete. Ranked fights are automated, and recorded ranked outcomes change its Elo. More training does not guarantee a stronger commander against every opponent or a higher place in Rankings.