Teaching the AI to Break Through

Post news Report RSS Teaching the AI to Break Through

TinyGenerals is a free, browser-based, turn-based hex wargame inspired by Advance Wars and Weewar. Capture buildings, build your army, and outmaneuver friends or AI. Play asynchronously with up to eight players, choose between two factions, and create your own battlefields with the built-in map editor.

Posted by nbwdesigns on Sep 11th, 2026

One TinyGenerals test battle lasted more than 3,000 turns. One side had a huge artillery force. The other kept feeding a few tanks into its guns. The defenders piled up kills, but never turned their advantage into a decisive attack.

TinyGenerals is a turn-based hex wargame about capturing buildings, producing units, and outmaneuvering another general. That endless battle exposed a gap between choosing a good individual action and using an army effectively.

Fixing the obvious mistakes

Two production bugs are fixed in the development version. A bot could recognise that it needed infantry to capture buildings, then buy another combat vehicle instead. Another rule stopped production when its army was overwhelmingly large—even if none of those units could capture anything.

We also fixed planning memory: one bot’s turn could overwrite another bot’s expansion assignments. Keeping a plan between turns is essential if units are going to achieve anything beyond their next move.

Giving the front room to move

The map had open space. The problem was which moves the AI would accept.

Units normally insist on getting closer to their target. In a crowded formation, that can leave a tank refusing to step sideways while infantry waits behind it. Moving aside would help the army, even though it would not shorten the tank’s own route.

We’re testing limited sideways moves, including a version that checks whether a specific friendly unit can advance into the freed space. These experiments reduced congestion and enabled more forward movement in the test situation. They remain disabled by default: moving the army more hasn't consistently made it fight better.

Firepower needs a follow-through

We’re also testing the balance between artillery and tanks. Another gun can look like a good purchase when the army already lacks units that can push forward and take defended ground.

In one experiment, more tanks meant more captures and fewer losses, but less damage dealt. Results on other maps were mixed. A strong kill count alone cannot tell us whether an army is breaking the defence or simply destroying replacements forever.

To judge these changes, we can now save a battle situation and compare two versions of the bot from the same starting point. That helps us see why units stop, what production choices change, and whether those changes actually help the attack.

The goal is an opponent that can bring its forces forward and exploit an opening. Coordinated attacks and recognizing opportunities to flank are the longer-term direction; the work now is making those plans possible.

Read the full devblog for more detail on the experiments →