Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Framework for Effective AI Recommendations in Cyber-Physical-Human Systems

Aditya Dave, Heeseung Bang, Andreas A. Malikopoulos

TL;DR

This work addresses AI-assisted decision-making in cyber-physical-human systems where humans may reinterpret or disregard AI recommendations. It establishes a principled $POMDP$-based framework and an information-state decomposition that yields optimal AI recommendations when the human model is known. To handle model uncertainty, the authors introduce an approximate human model (AHM) with provable bounds on the optimality gap, and propose data-driven methods to construct AHMs via supervised learning. A numerical example demonstrates the benefit of incorporating AHMs in partially observed settings, showing improved performance over naive or non-informed strategies. The framework enables principled design and learning of human-aware AI recommendations for complex CPHS applications.

Abstract

Many cyber-physical-human systems (CPHS) involve a human decision-maker who may receive recommendations from an artificial intelligence (AI) platform while holding the ultimate responsibility of making decisions. In such CPHS applications, the human decision-maker may depart from an optimal recommended decision and instead implement a different one for various reasons. In this letter, we develop a rigorous framework to overcome this challenge. In our framework, we consider that humans may deviate from AI recommendations as they perceive and interpret the system's state in a different way than the AI platform. We establish the structural properties of optimal recommendation strategies and develop an approximate human model (AHM) used by the AI. We provide theoretical bounds on the optimality gap that arises from an AHM and illustrate the efficacy of our results in a numerical example.

A Framework for Effective AI Recommendations in Cyber-Physical-Human Systems

TL;DR

This work addresses AI-assisted decision-making in cyber-physical-human systems where humans may reinterpret or disregard AI recommendations. It establishes a principled -based framework and an information-state decomposition that yields optimal AI recommendations when the human model is known. To handle model uncertainty, the authors introduce an approximate human model (AHM) with provable bounds on the optimality gap, and propose data-driven methods to construct AHMs via supervised learning. A numerical example demonstrates the benefit of incorporating AHMs in partially observed settings, showing improved performance over naive or non-informed strategies. The framework enables principled design and learning of human-aware AI recommendations for complex CPHS applications.

Abstract

Many cyber-physical-human systems (CPHS) involve a human decision-maker who may receive recommendations from an artificial intelligence (AI) platform while holding the ultimate responsibility of making decisions. In such CPHS applications, the human decision-maker may depart from an optimal recommended decision and instead implement a different one for various reasons. In this letter, we develop a rigorous framework to overcome this challenge. In our framework, we consider that humans may deviate from AI recommendations as they perceive and interpret the system's state in a different way than the AI platform. We establish the structural properties of optimal recommendation strategies and develop an approximate human model (AHM) used by the AI. We provide theoretical bounds on the optimality gap that arises from an AHM and illustrate the efficacy of our results in a numerical example.
Paper Structure (8 sections, 8 theorems, 11 equations, 3 figures)

This paper contains 8 sections, 8 theorems, 11 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Recommendation Framework
  4. Optimal recommendation strategies
  5. Approximate human model
  6. Constructing an approximate human model
  7. Numerical Example
  8. Concluding Remarks

Key Result

Lemma 1

Given a human model, Problem problem_1 is equivalent to computing the optimal strategy in a POMDP with state $(X_t, S_t) \in \mathcal{X} \times \mathcal{S}$, input $U_t^{\text{ai}} \in \mathcal{U}$, observation $(Y_t, U_{t-1}^{\text{h}}) \in \mathcal{Y} \times \mathcal{U}$, and reward $R_t \in [r^{\

Figures (3)

  • Figure 1: Control loop of the recommendation problem.
  • Figure 2: System model for the machine.
  • Figure 3: Rewards obtained using different strategies.

Theorems & Definitions (21)

  • Lemma 1
  • proof
  • Lemma 2
  • proof
  • Lemma 3
  • proof
  • Lemma 4
  • proof
  • Lemma 5
  • proof
  • ...and 11 more