For some, training a neural network might seem daunting - especially when it seems to not always learn what you want it to learn. Although they are very powerful models, it might not be the easiest start to AI Arena. As such, we created a much simpler model that’s easier to train!

At a high level, you can imagine this model as a table with multiple “cells”. Each of these cells stores information about what you showed the model in training. Before walking through how each of the cells are divided throughout our various “buckets” (which is another word for how we group things), we want to note a few things:

• When getting to the end of a sub-bucket chain, you will see “cell N”, where N represents the cell in which that specific sub-bucket belongs.
• If a sub-bucket has “(Inverted)” in it, that means that all cells within the sub-bucket will have auto-inverting applied to them. All inverted sub-buckets have the default as your fighter being on the left side.
• If a cell has “(Perception Delay)” in it, it means that cell has a delay applied to it, it delays the registering of the opponent’s attack so your AI does not immediately react, which is similar to how it works for humans.
• “(Rough Range)”, is a new term for projectile-related cells, when a projectile trajectory would be in-line traveling towards the opponent given some buffer range, this is considered within Rough Range. This buffer is to prevent the cells from acting as an aimbot.

Buckets

Recovery (you off-stage)

Vertical Combat

Advantage (invert)

Disadvantage (invert)

Neutral (invert)