Explainable and Interpretable Reinforcement Learning for Robotics (Synthesis Lectures on Artificial Intelligence and Machine Learning)

✍ Scribed by Aaron M. Roth, Dinesh Manocha, Ram D. Sriram, Elham Tabassi

• Publisher: Springer
• Year: 2024
• Tongue: English
• Leaves: 123
• Edition: 2024
• Category: Library

✦ Synopsis

This book surveys the state of the art in explainable and interpretable reinforcement learning (RL) as relevant for robotics. While RL in general has grown in popularity and been applied to increasingly complex problems, several challenges have impeded the real-world adoption of RL algorithms for robotics and related areas. These include difficulties in preventing safety constraints from being violated and the issues faced by systems operators who desire explainable policies and actions. Robotics applications present a unique set of considerations and result in a number of opportunities related to their physical, real-world sensory input and interactions.

The authors consider classification techniques used in past surveys and papers and attempt to unify terminology across the field. The book provides an in-depth exploration of 12 attributes that can be used to classify explainable/interpretable techniques. These include whether the RL method is model-agnostic or model-specific, self-explainable or post-hoc, as well as additional analysis of the attributes of scope, when-produced, format, knowledge limits, explanation accuracy, audience, predictability, legibility, readability, and reactivity. The book is organized around a discussion of these methods broken down into 42 categories and subcategories, where each category can be classified according to some of the attributes. The authors close by identifying gaps in the current research and highlighting areas for future investigation.

✦ Table of Contents

Acknowledgements
Contents
About the Authors
1 Introduction
1.1 Motivation
1.2 Background
1.2.1 Reinforcement Learning
1.2.2 Reinforcement Learning Versus Supervised, Unsupervised, and Imitation Learning
1.2.3 Explainable Artificial Intelligence (XAI) and Explainable Reinforcement Learning (XRL)
1.2.4 Explainable Robotics (X-Robotics)
1.3 Selection Criteria
2 Classification System
[DELETE]
2.1 Prior Surveys on XRL or X-Robotics
2.2 Existing Classification Terminology for XRL or X-Robotics
2.3 The Attributes of our Classification System
2.3.1 Hard Attributes
2.3.2 Soft Attributes (General)
2.3.3 Soft Attributes (Robot-Specific)
3 Explainable Methods Organized by Category
[DELETE]
3.1 Decision Tree
3.1.1 Single Decision Tree
3.1.2 Single Altered Decision Tree
3.1.3 Multiple or Combined Decision Trees
3.2 Saliency Maps
3.2.1 Post-Hoc Saliency Maps via Backpropagation
3.2.2 Intrinsic Saliency Maps
3.2.3 Post-Hoc Saliency Maps via Input Perturbation
3.3 Counterfactuals/Counterexamples
3.3.1 Counterfactual by Input Perturbation or Extra Information
3.3.2 Counterfactual by Model Checking
3.4 State Transformation
3.4.1 Dimension Reduction
3.4.2 Meaningful Representation Learning
3.5 Observation Based Methods
3.5.1 Observation Analysis: Frequency or Statistical Techniques for Policy Understanding
3.5.2 Observation Analysis: Human Communicative Trajectories for Goal Understanding
3.5.3 Observation Analysis: A/B Testing
3.5.4 Training Data Observation Analysis
3.5.5 Interrogative Observation Analysis
3.6 Custom Domain Language
3.7 Constrained Learning
3.8 Constrained Execution
3.9 Hierarchical
3.9.1 Hierarchical Skills or Goals
3.9.2 Primitive Generation
3.10 Machine-to-Human Templates
3.10.1 Model-to-Text or Policy-to-Text Templates
3.10.2 Query-Based NLP Templates
3.11 Model Reconciliation
3.11.1 Certain Model Reconciliation
3.11.2 Uncertain Model Reconciliation
3.12 Causal Methods
3.13 Reward Decomposition
3.13.1 Standard Reward Decomposition Methods
3.13.2 Model Uncertainty Reward Decomposition
3.14 Visualizations
3.15 Instruction Following
3.16 Symbolic Methods
3.16.1 Symbolic Transformation
3.16.2 Symbolic Reward
3.16.3 Symbolic Learning
3.17 Legibility or Readability
4 Key Considerations and Resources
[DELETE]
4.1 General Discussion
4.2 Limitations of Some Methods in this Survey
4.3 Human-Robot Interaction Considerations
4.4 Legibility and Readability
4.5 AI Safety
4.6 Environments
5 Opportunities, Challenges, and Future Directions
[DELETE]
5.1 Opportunities, Challenges, and Future Directions
5.2 Conclusion
References
Index