Shrink the Game
The action space in SC2 is absurd. Hundreds of actions, dozens of units, fog of war, real-time pressure, and rewards that show up ten minutes later. If you hand all of that to an agent and say only one thing, win, it has no chance. There is no signal and nothing to latch onto.
So you shrink the world. One map. One scenario. One matchup. Sometimes even one unit. Makes traction way more possible.
Even AlphaStar controlled its variables. No reason you shouldn’t too.
There’s More Than One Way
When someone hears ML they think, neural networks, but lets slow the roll there because. That’s dialing the difficulty crank to 100.
Almost every serious SC2 AI project is hybrid. You hardcode the boring, brittle stuff. You filter actions, add basic rules, give the bot fallback behaviors, and stabilize the early game with simple logic. Then ML handles improvement inside that structure.
As tempting as it might seem, you also don’t need to chase fancy unsupervised setups. In SC2, the game naturally lends itself to reinforcement learning and supervised learning from replays: clear actions, clear rewards, and tons of human games to mine.
You cannot let an agent choose from 800 actions every frame. You have to reduce the action space, bundle decisions, and gate choices. Without priors, the bot never learns anything useful.
Scope is Your Best Friend
This is the lesson almost nobody believes at first.
The ML that actually works in SC2 is always scoped to one job. Zerg rush detection. Range stutter stepping. Strategy choosing. Small, tightly defined problems with immediate feedback.
Trying to learn the whole game at once is a good way to end up pulling your hair out. This will also reduce the time and resources ($) to train it, especially when you’re doing reinforcement learning loops.
My take
ML feels “impossible” when you think you are teaching StarCraft. You are not. You are teaching one slice of it. Shrink the world, add structure, and train tiny problems. Do that and ML starts becoming a real tool you can use.
I spoke to some bot makers across different games (including someone who worked on Alphastar) to get their takes on how you build an ML bot without breaking the bank, tons of insight
Watch it
