Cognitive Science as a Source of Forward and Inverse Models of Human Decisions for Robotics and Control

Mark K. Ho, Thomas L. Griffiths

Introduction

As autonomous systems transition from controlled industrial environments to the dynamic complexity of human-centric spaces, they encounter a fundamental "human problem." Humans are not just obstacles to be avoided; they are intentional agents whose behavior is shaped by internal states and who react to the robot's presence. In their review, Mark K. Ho and Thomas L. Griffiths argue that cognitive science provides the essential computational tools—forward and inverse models—to bridge this gap.

The Core Challenge: Prediction and Inference

For a robot to interact effectively with a human, it must solve two primary computational problems. First, it requires Forward Models to predict human actions based on their likely goals. Second, it needs Inverse Models to infer a human’s hidden mental states, such as beliefs and desires, from their observed actions.

Cognitive components like beliefs and desires are mapped to decisions through forward models (prediction) and inverse models (inference).

Historically, robotics has relied on two extremes: perfectly rational models that generalize well but fail to capture human nuance, or specific heuristic lists that are accurate but brittle. Cognitive science offers a middle path by formalizing the internal "cloud" of human cognition—including biases, habits, and planning—into actionable models.

Computational Frameworks for Human Behavior

The authors highlight several frameworks that move beyond simple rationality:

  • Expected Utility Theory (EUT): The baseline model where agents maximize expected rewards.
  • Resource Rationality: This framework posits that human "biases" are actually optimal strategies given finite computational resources. It uses meta-level Markov Decision Processes (MDPs) to model the cost of thinking itself.
  • Theory of Mind (ToM): Bayesian models that allow machines to "parse" human behavior into underlying mental states.
  • Rational Speech Act (RSA): A model of communication where agents recursively reason about each other's intentions to achieve legible interaction.
Evolution of decision models from classical expected utility (A) and heuristics (B) to neural networks (C) and resource rationality (D).

Empirical Evidence and Datasets

The paper validates these frameworks through diverse datasets:

  • Risky Choice Datasets: Comparing small-scale psychological experiments (90 pairs) with massive datasets (10,000+ pairs), researchers found that while neural networks eventually outperform classic theories, they require significantly more data to "discover" the same psychological principles.
  • The Moral Machine Project: Utilizing over 10 million decisions, researchers trained interpretable models to understand human ethical preferences in autonomous vehicle scenarios.
  • Grid Navigation: Experiments in "Showing vs. Doing" demonstrate that humans intentionally alter their behavior to be more communicative (legible) when they know they are being observed.

Key Insights: Beyond Rationality

A central theme is the Accuracy-Generalizability Trade-off. While black-box machine learning (ML) models can be highly accurate with enough data, theory-driven cognitive models generalize better to novel situations. The authors propose Hybrid Modeling, where ML "critiques" and refines interpretable psychological models.

Furthermore, the research emphasizes that human "errors" are often Bounded Optimality. For instance, focusing on extreme events is an efficient way to minimize estimation variance when mental sampling is limited. Finally, the distinction between "Showing" and "Doing" is crucial: robots should not just mimic human behavior but understand when humans are attempting to teach them.

Future Directions

Despite progress, several challenges remain. Inverse Resource Rationality—how humans perceive the cognitive limits of others—is still poorly understood. Additionally, scaling Theory of Mind for real-time robotic applications requires more computationally tractable implementations. Finally, moving from dyadic interactions to broader social norms and joint intentions is the next frontier for creating truly social autonomous agents.

Summary updated

Abstract

Those designing autonomous systems that interact with humans will invariably face questions about how humans think and make decisions. Fortunately, computational cognitive science offers insight into human decision-making using tools that will be familiar to those with backgrounds in optimization and control (e.g., probability theory, statistical machine learning, and reinforcement learning). Here, we review some of this work, focusing on how cognitive science can provide forward models of human decision-making and inverse models of how humans think about others’ decision-making. We highlight relevant recent developments, including approaches that synthesize black box and theory-driven modeling, accounts that recast heuristics and biases as forms of bounded optimality, and models that characterize human theory of mind and communication in decision-theoretic terms. In doing so, we aim to provide readers with a glimpse of the range of frameworks, methodologies, and actionable insights that lie at the intersection of cognitive science and control research. Keyword(s): cognitive science, decision-making, psychology, robotics, theory of mind
BibTeX
@article{ho2022cognitive,
  author = {Mark K. Ho and Thomas L. Griffiths},
  title = {{Cognitive Science as a Source of Forward and Inverse Models of Human Decisions for Robotics and Control}},
  journal = {Annu. Rev. Control. Robotics Auton. Syst.},
  year = {2022},
  doi = {10.1146/annurev-control-042920-015547},
}
Figure