Self reward design with fine-grained interpretability.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) exhibit a "black-box" nature, raising concerns about transparency and fairness.
• Deep Reinforcement Learning (DRL), reliant on DNNs, inherits these transparency and fairness challenges.

Purpose of the Study:

• To propose a novel framework, Self Reward Design (SRD), for creating interpretable deep reinforcement learning models.
• To address the black-box nature of DNNs in DRL by designing neural networks with human-understandable concepts.

Main Methods:

• Introduced the Self Reward Design (SRD) framework, inspired by Inverse Reward Design.
• Designed neural networks bottom-up, ensuring each neuron/layer has a meaningful, human-understandable utility.
• Demonstrated SRD's capability to solve RL problems like lavaland and MuJoCo with minimal parameters.

Main Results:

• SRD allows for problem-solving through deliberate design, albeit imperfectly.
• The interpretable SRD models can be optimized using standard DNN techniques.
• Successfully applied SRD to a fish sale auction example, showcasing its utility in scenarios requiring semantic-based decisions.

Conclusions:

• SRD provides a method to circumvent transparency and fairness issues in DRL.
• Interpretable neural network design enables human-understandable decision-making in complex AI systems.
• SRD facilitates the development of more trustworthy and explainable AI solutions.