There are many architectures used in deep learning - you can check out a few here. For AI Arena, we focus on feedforward neural networks. In general they have fewer parameters than convolutional neural networks (CNNs) or recurrent neural networks (RNNs).

We performed feature engineering to reduce the size of the network needed to train a good policy. Below we outline all of the features groups in the advanced configuration and the features that are contained within those groups.

Feature Groups

• Raycast Platform Distance (16 features)
• Raycast Platform Type (16 features)
• Raycast Projectile Distance (16 features)
• Raycast Projectile on Target (16 features)
• Raycast Projectile Element (16 features)
• Action Representation (18 features)
• Your Positioning (2 features)
• Opponent Positioning (2 features)
• Direction (2 features)
• Distance to Opponent (1 feature)
• Angle to Opponent (2 features)
• Stocks (2 features)
• Percentage (2 features)
• Elemental Gauge (2 features)
• Elemental Advantage (1 feature)