Nets

The nets module contains the core neural network building blocks used across ObjectRL’s algorithms. It is organized to facilitate easy customization of network architectures.

Module Structure

• layers: Contains reusable neural network layers and utilities such as Bayesian layers, normalization layers, and other custom components that can be used by actor and critic networks.

• actor_nets.py: Defines policy network architectures responsible for selecting actions given states. These networks encapsulate the actor’s behavior and can be customized or extended for new algorithms.

• critic_nets.py: Defines value or Q-function network architectures responsible for estimating expected returns. Critic networks support various architectures, including standard MLPs and Bayesian models.

Purpose and Usage

This organization allows you to:

• Easily swap or extend network architectures for actors and critics without modifying the core algorithm logic.

• Reuse custom layers across different network types.

• Implement novel network designs to experiment with uncertainty estimation, exploration bonuses, or other advanced features.