Explainable Methods Organized by Category

This chapter describes specific state-of-the-art methods for XRL-Robotics, organized by category. Each category and subcategory are described according to some of the attributes in our classification system, summarized in Tables 3.1, 3.2, 3.3 and 3.4. The methods in each category are described and discussed. The categories are Decision Tree, Single Decision Tree, Single Altered Decision Tree, Multiple or Combined Decision Trees, Saliency Maps, Post-Hoc Saliency Maps via Backpropagation, Intrinsic Saliency Maps, Post-Hoc Saliency Maps via Input Perturbation, Counterfactuals/Counterexamples, Counterfactual by Input Perturbation or Extra Information, Counterfactual by Model Checking, State Transformation, Dimension Reduction, Meaningful Representation Learning, Observation Based Methods, Observation Analysis: Frequency or Statistical Techniques, for Policy Understanding, Observation Analysis: Human Communicative Trajectories, for Goal Understanding, Observation Analysis: A/B Testing, Training Data Observation Analysis, Interrogative Observation Analysis, Custom Domain Language, Constrained Learning, Constrained Execution, Hierarchical, Hierarchical Skills or Goals, Primitive Generation, Machine-to-Human Templates, Model-to-Text or Policy-to-Text Templates, Query-based NLP Templates, Model Reconciliation, Certain Model Reconciliation, Uncertain Model Reconciliation, Causal Methods, Reward Decomposition, Standard Reward Decomposition Methods, Model Uncertainty Reward Decomposition, Visualizations, Instruction Following, Symbolic Methods, Symbolic Transformation, Symbolic Reward, Symbolic Learning, and Legibility or Readability.