Overview
For AI Arena we hand-crafted the variables because we want to initially constrain the size of neural networks used. As a result, researchers do not need to worry about making massive CNNs, which can be computationally intensive.
One of our core values is to provide equal opportunity for all - we want to reward users for being able to train their models better, not for having more resources.
Traditional Fighter State Space
In the current version of the game - traditional fighting game - the state space is comprised of 9 features:
Platform Fighter State Space
In the next version of the game - platform fighting game - the state space will increase approximately 10x because there is significantly more mobility and things a model must consider. We will release more details about this once the new game engine is ready to be deployed.
State | Calculation | Range |
Relative Distance (RD) | location1 - location2 | [-1, 1] |
You Facing Opponent | -sign(RD) * direction1 | {0, 1} |
Opponent Facing You | sign(RD) * direction2 | {0, 1} |
Your Health | health1 | (0, 1] |
Opponent Health | health2 | (0, 1] |
Relative Power | power1 - power2 | (-1, 1) |
Relative Speed | speed1 - speed2 | (-1, 1) |
Relative Defense | defense1 - defense2 | (-1, 1) |
Relative Accuracy | accuracy1 - accuracy2 | (-1, 1) |
