PokAImon Red ROM

So, the basic idea behind PokAImon Red is simple. I wanted to find out how far the original Pokémon Red battle engine could be pushed without replacing the game around it. The story, region, progression, and general Gen 1 identity are still Pokémon Red. What changes is the opponent sitting on the other side of the battle.

The original trainer AI is mostly a short collection of rules. It can prefer certain move effects, but it does not really understand the complete position in front of it. It regularly misses obvious finishes, wastes status moves, heals at strange times, and makes decisions without a useful picture of damage, status, stat stages, or the team behind the active Pokémon. PokAImon Red rebuilds that decision process around a compact neural policy which runs inside the ROM on actual Game Boy hardware.

The important part is that this is not a modern program controlling an emulator from the outside. Training happens on a computer beforehand, but the finished network, its weights, its feature extraction, and the complete inference routine are stored in the cartridge image. When an enemy trainer chooses an action, the Game Boy itself evaluates the current battle and produces the decision.

How the AI works

The pipeline starts outside the game by simulating battles and using Monte Carlo Tree Search as the teacher. Each decision is recorded with the state that produced it, including move damage, accuracy, knockout chances, HP, status conditions, remaining sleep turns, stat stages, move effects, and the living Pokémon available on the bench. Labels which point at actions that provably do nothing are removed, and available finishing moves are reinforced before training.

From there, a small 60 to 40 to 6 neural network is trained. The 60 values describe the battle, the 40 hidden units learn the useful combinations between them, and the six outputs represent four moves and two possible switches. Once training is complete, the weights are reduced to signed 8 bit integers and exported directly into assembly tables. Inside the ROM, the network uses fixed point arithmetic rather than floating point.

The inference engine performs both layers with a lookup table multiplication method, keeps the accumulators within bounds, applies the hidden activation, masks illegal actions, and selects the highest legal score. The buffers are placed in safe cartridge SRAM scratch space because the original game has almost no free working RAM left. In short, the expensive learning happens once on modern hardware, while every actual decision happens locally inside the Game Boy.