AI-ML for Decision and Risk Analysis: Challenges and Opportunities for Normative Decision Theory

Preface
Acknowledgments
Contents
Part I: Received Wisdom
	Chapter 1: Rational Decision and Risk Analysis and Irrational Human Behavior
		Introduction
		Extending Classical Decision Analysis (DA) with AI/ML Ideas
		Decision Analysis for Realistically Irrational People
		Two Decision Pathways: Emotions and Reasoning Guide Decisions
		We All Make Predictable Mistakes
		Marketers, Politicians, Journalists, and Others Exploit Our Systematic Mistakes
		People Respond to Incentives and Influences in Groups, Organizations, and Markets
		Moral Psychology and Norms Improve Cooperative Risk Management
		We Can Learn to Do Better
		The Rise of Behavioral Economics
		Beyond Behavioral Economics and Rational Choice: Behaving Better in a Risky World
		Review of Behave: The Biology of Humans at Our Best and Worst
		12 Rules for Life: An Antidote to Chaos
		Comments on Behave and 12 Rules
		Review of Morality: Restoring the Common Good in Divided Times
		References
	Chapter 2: Data Analytics and Modeling for Improving Decisions
		Introduction
		Forming More Accurate Beliefs: Superforecasting
		Learning About the World Through Data Analysis: The Art of Statistics
		Using Models to Interpret Data: The Model Thinker
			Overview of Contents
		Comments on The Model Thinker
		Responding to Change Strategically: Setting Goals and Acting Under Uncertainty
		Using Data to Discover What Works in Disrupting Poverty
			Conceptual Framework: Uncertain Risks and Rewards and Poverty Traps
			Extending and Applying the Framework: How Risk and Perceptions Strengthen Poverty Traps
			Escaping Poverty Traps: Weakly Held Beliefs and Credible Communication
			How Can Analysis Help Reduce Health Risks and Poverty?
		References
	Chapter 3: Natural, Artificial, and Social Intelligence for Decision-Making
		Introduction
		Biological Foundations of Thought: Cognitive Neuroscience
		Thinking and Reasoning
		Computational Models of Nondeliberative Thought: Deep Learning (MIT Press, 2019)
		Computational Models of Deliberative Thought: Artificial Intelligence: A Very Short Introduction (Oxford University Press, 201...
		Communities of Knowledge
		Factfulness: Ten Reasons We´re Wrong About the World-and Why Things Are Better Than You Think
		Aligning AI-ML and Human Values: The Alignment Problem
		Conclusions
		References
Part II: Fundamental Challenges for Practical Decision Theory
	Chapter 4: Answerable and Unanswerable Questions in Decision and Risk Analysis
		Introduction: Risk Analysis Questions
		Some Models and Methods for Answering Risk Analysis Questions
			The Simplest Causal Models: Decision Tables and Decision Trees
			Fault Trees, Event Trees, Bayesian Networks (BNs) and Influence Diagrams (IDs)
			Markov Decision Processes (MDPs) and Reinforcement Learning (RL)
			Simulation-Optimization for Continuous, Discrete-Event, and Hybrid Simulation Models
			Response Surface Methodology
			Adaptive and Robust Control
			Distributed and Hierarchical Control of Uncertain and Non-stationary Systems
			Decentralized Multi-agent Control: POMDP, decPOMD, SMDP, and POSMDP Models
			Agent-Based Models (ABMs) and Cellular Automata (CA)
			Game Theory Models and Adversarial Risks
		Undecidability: Not All Risk Analysis Questions Can Be Answered in All Causal Models
		Undecidable Questions in Hazard Identification and Probabilistic Risk Assessment
		Undecidable Questions in Risk Management
			Control of Deterministic Dynamic Systems
			Risk Management of Uncertain Systems Modeled by POMDPs
			Monitoring and Probabilistic Fault Diagnosis in Partially Observable Systems: Timeliness vs. Accuracy
			Guaranteeing Timely Resolution of Tasks with Uncertain Completion Times
			Multi-agent Team Control
			Risk Management with Intelligent Adversaries: Game Theory Models
		Learning Causal Models from Data
		Responding to the Intractability of Risk Analysis
			Risk Analysis for Restricted Systems: Complexity-Tractability Trade-Offs
			Design of Resilient Systems for More Tractable Risk Management
		Open-World vs. Closed-World Risks
		Artificial Intelligence (AI) Methods for Coping with Open-World Novelty and Risk
			Behavior Trees (BTs) Enable Quick Responses to Unexpected Events While Maintaining Multiple Goals
			Integrated Machine Learning and Probabilistic Planning for Open Worlds
			Anomaly Detection Helps Focus Attention When Needed
			Summary: AI Capabilities for Dealing with Open-World Risks and Uncertainties
		Discussion and Conclusions: Thriving with a Mix of Answerable and Unanswerable Questions
		References
	Chapter 5: Decision Theory Challenges for Catastrophic Risks and Community Resilience
		Introduction
		Challenges of Rare Catastrophic Events to Traditional Analytical Methods
			Unpredictability of Catastrophes in Physical, Biological, and Social Systems
			Example: Self-Organizing Criticality Makes the Size and Timing of System Responses Unpredictable
		Example: Poisson Arrival of Rare Catastrophic Events
		Example: Unpredictability in Deterministic Physical and Ecological Models
		Example: Deterministic Chaos Limits Possible Forecast Horizons
		Decision Analysis Can Omit Crucial Details in Describing Catastrophes
			Example: Risk Curves for Frequency and Severity Do Not Show Risk Equity
		Emergent Precautionary Choice Behaviors Can Be Incoherent and Unpredictable
			Example: Coherent Individual Preferences Can Create Incoherent Group Choices
			Example: Dynamic Inconsistency of Majority Preferences for Costly Precautions
		Challenges to Normative Group Decision Theory for Risk Management
			Example: Aggregating Individual Beliefs Can Lead to Group Risk Management Decisions that No One Likes
		Toward a New Foundation for Disaster Risk Management: Building Disaster-Resilient Communities
			Example: Resilient Response to the North Sea Flood of 1953
		Bistability and the Evolution and Collapse of Social Cooperation
		Summary and Conclusions
		References
	Chapter 6: Learning Aversion in Benefit-Cost Analysis with Uncertainty
		Introduction: Benefit-Cost Analysis (BCA) Fundamentals
		Aspirations and Benefits of BCA
			Example: Majority Rule Without BCA Can Yield Predictably Regrettable Collective Choices
		Limitations of BCA for Purely Rational People, Homo Economicus
			Example: Pareto-Inefficiency of BCA with Disagreements About Probabilities
			Example: Impossibility of Pareto-Efficient Choices with Sequential Selection
		How Real People Evaluate and Choose Among Alternatives
		Learning Aversion and Other Decision Biases Inflate WTP for Uncertain Benefits
			Example: Overconfident Estimation of Health Benefits from Clean Air Regulations
		Assuming No Risk Aversion Inflates the Estimated Value of Public Projects with Uncertain Benefits
			Example: Information Externalities and Learning Aversion in Clinical Trials
			Example: Desirable Interventions with Uncertain Benefits Become Undesirable When They Are Scaled Up
		Doing Better: Using Predictable Rational Regret to Improve BCA
			Example: Rational vs. Irrational Regret
		Conclusions
		References
Part III: Ways Forward
	Chapter 7: Addressing Wicked Problems and Deep Uncertainties in Risk Analysis
		Introduction: How to Make Good Decisions with Deep Uncertainties?
		Principles and Challenges for Coping with Deep Uncertainty
		Point of Departure: Subjective Expected Utility (SEU) Decision Theory
		Four Major Obstacles to Applying SEU to Risk Management with Model Uncertainty
		Ten Tools of Robust Risk Analysis for Coping with Deep Uncertainty
		Use Multiple Models and Relevant Data to Improve Decisions
		Robust Decisions with Model Ensembles
			Example: Robust Decisions with Model Uncertainty
			Example: Robustness, Multiple Models, Ambiguous Probabilities, and Multiple Priors
			Example: Robust Optimization and Uncertainty Sets Using Coherent Risk Measures
		Averaging Forecasts
		Resampling Data Allows Robust Statistical Inferences in Spite of Model Uncertainty
		Adaptive Sampling and Modeling: Boosting
		Bayesian Model Averaging (BMA) for Statistical Estimation with Relevant Data But Model Uncertainty
		Learning How to Make Low-Regret Decisions
			Example: Learning Low-Regret Decision Rules with Unknown Model Probabilities
		Reinforcement Learning of Low-Regret Risk Management Policies for Uncertain Dynamic Systems
		Example: Reinforcement Learning of Robust Low-Regret Decision Rules
		Example: Model-Free Learning of Optimal Stimulus-Response Decision Rules
		Applying the Tools: Accomplishments and Ongoing Challenges for Managing Risks with Deep Uncertainty
			Planning for Climate Change and Reducing Energy Waste
			Sustainably Managing Renewable Resources and Protecting Ecosystems
			Managing Disease Risks
			Maintaining Reliable Network Infrastructure Service Despite Disruptions
			Adversarial Risks and Risks from Intelligent Agents
		Conclusions
		References
	Chapter 8: Muddling-Through and Deep Learning for Bureaucratic Decision-Making
		Introduction: Traditional Benefit-Cost Analysis (BCA) and Decision Analysis
		Developments in Rational-Comprehensive Models of Decision-Making
		Modern Algorithms for Single- and Multi-Agent Decision-Making
		Discussion: Implications of Advances in Rational-Comprehensive Decision Theory for Muddling Through
		Conclusions
		References
	Chapter 9: Causally Explainable Decision Recommendations Using Causal Artificial Intelligence
		Introduction: Creating More Trustworthy AI/ML for Acting Under Risk and Uncertainty
		The Structure of Traditional Statistical Explanations
		The Structure of Explanations in Causal Bayesian Networks (BNs)
			Explaining Direct, Indirect (Mediated), and Total Effects
			Conditional Independence and Potential Causality in BNs
		Causal Discovery for Predictive, Interventional, and Mechanistic Causation
		Knowledge-Based Constraints Help to Orient Arrows to Reflect Causal Interpretations
		Structure of Most Probable Explanations (MPEs) in Bayesian Networks
		Explaining Predictions for Effects of Interventions: Adjustment Sets, Causal Partial Dependence Plots (PDPs) and Accumulated L...
		Structure of Explanations for Decision and Policy Recommendations in Influence Diagrams (IDs): Maximizing Expected Utility wit...
		Structure of Explanations for Decision Recommendations Based on Monte Carlo Tree Search (MCTS) and Causal Simulation
		Structure of Explanations for Decision Recommendations Based on Reinforcement Learning (RL) with Initially Unknown or Uncertai...
		Limitations and Failures of Causally Explainable Decisions for Dynamic Systems
		Discussion: Explaining CAI Decision Recommendations
		Applying CAI Principles to Explain Decision Recommendations
		Conclusions: Explaining Recommended Decisions in Causal AI
		References
Part IV: Public Health Applications
	Chapter 10: Re-Assessing Human Mortality Risks Attributed to Agricultural Air Pollution: Insights from Causal Artificial Intel...
		Introduction
		Limitations of Current Qualitative Hazard Identification for PM2.5-Mediated Health Effects of NH3 Emissions
			Burden of Disease Calculations
			Deaths Attributed to Air Pollution Do Not Refute Non-Causal Explanations
			Attributed Risks Do Not Predict Effects of Interventions
			Attributed-Mortality Calculations Do Not Reveal Effects on Mortality of Reducing Air Pollution
		Intervention Studies Do Not Support Attributed Risks Based on Associations
		Some Limitations of the Quantitative Risk Assessment
		Considerations from Causal Artificial Intelligence (CAI) and Machine Learning
			Conditional Independence Analysis
			Granger Causality Analysis
			Non-Parametric Analysis of Heterogeneous Exposure Effects
			Invariant Causal Prediction and Transportability of Estimated Causal Effects
			Probabilistic Causal Network Modeling of Multiple Interrelated Variables
		Discussion and Conclusions
		References
	Chapter 11: Toward More Practical Causal Epidemiology and Health Risk Assessment Using Causal Artificial Intelligence
		Introduction: Why Are Better Causal Methods Needed in Applied Epidemiology?
		Thinking About Causation: A Thought Experiment with Dominos
			Different Concepts of Causation
			Clarifying Causation with Conditional Probability Networks
			Pivoting Epidemiology from Attribution to Causal Prediction
		CAI Conceptual Framework: Qualitative Structure of Causal Networks of Probabilistic Causal Mechanisms
		Practical Algorithms for Quantitative Causal Inference and Prediction with Realistically Imperfect Data
		Implications of CAI for Calculating and Interpreting Preventable Fractions
			Seeing Is Not Doing
			Ambiguity of Counterfactuals for PAFs
		Discussion and Conclusions: Toward Pragmatic Causal PAF Calculations
		References
	Chapter 12: Clarifying the Meaning of Exposure-Response Curves with Causal AI and ML
		Introduction: What Does an Exposure-Response Curve Mean?
		Exposure-Response Regression Curves Describe Responses at Different Observed Exposures
			A Point of Departure: Correlation vs. Causality
			Assumption-Dependent Causal Interpretations of Exposure-Response Regression Models
			Heterogeneity in Individual Risks
			Ambiguous Regression Coefficients: Inference vs. Intervention
		Logistic Regression vs. Non-Parametric Exposure-Response Curves
		Partial Dependence Plots (PDPs)
		Describing Interindividual Heterogeneity in Exposure-Response Functions: Individual Conditional Expectation (ICE) Plots
		Relevant and Data-Informed Counterfactuals: Two-Dimensional Partial Dependence Plots (2D-PDPs)
		Discussion and Conclusions: What Do We Want Exposure-Response Curves to Mean?
		References
	Chapter 13: Pushing Back on AI: A Dialogue with ChatGPT on Causal Inference in Epidemiology
		Introduction
		Dialogue with ChatGPT on Causal Interpretation of PM2.5-Mortality Associations
		Discussion
		Conclusion
		References
Index