The current stable and up to date simbot weights, the minecraft AI reinforcement learning player neural network. (Code soon to be publicly released.)
Public demo is on my friends minecraft server mc.nico.re (Plus runs on most up to date weights)
Model specs:
Layers:
- 731 (Input; blocks, entities, position, last eaten, health, and more)
- 2048 (Hidden layer)
- 1024 (Hidden layer)
- 512 (Hidden layer)
- 66 (Output; what action to do, eg. eat, fight, look, move...)
Total params: 4.155m.
Example capabilities:
- Eats food when low on hunger.
- Eats Golden apples when low on health.
- Sometimes flees, sometimes fights, based off free space and ammount of enemies.
- Can beat enemies
Known limitations:
- Randomly bumps into walls.
- Forgets things.
- Sometimes fights using random items instead of swords.
Reward system:
Rewarded: Eating, Damaging a enemy, sneaking near players (Just for fun).
Punished: Getting damaged, Dieing, Stuck in a loop, Stuck in one area, Clicking at nothing.
