Benjamin Van Roy: Opensource Code

Epistemic Neural Networks. This is a general interface for uncertainty estimation in deep learning. All existing approaches to uncertainty modeling, such as Bayesian neural networks (BNNs), can be expressed as ENNs. However, there are ENN architectures that can not be expressed as BNNs. This library provides interfaces and tools for the design and training of ENNs.

ENN Agent in Acme. This defines an interface for designing RL agents that combines ENNs and the Acme RL library.

The Neural Testbed. This is a system for evaluating the performance of epistemic neural networks, which are models that generate joint predictions over multiple outputs in response to multiple inputs. Included along with evaluation tools is a set of baseline agents.

Behavior Suite for Reinforcement Learning. This is system for evaluating the ability of reinforcement learning agents to meet simple challenges. The system aids in understanding strengths and weaknesses of an agent's behavior. Included along with this evaluation system is a set of baseline agents.