# Optimizing Risk-averse Human-AI Hybrid Teams

**Authors:** Andrew Fuchs, Andrea Passarella, Marco Conti  
**arXiv:** [2403.08386](https://arxiv.org/abs/2403.08386)  
**Date:** 2024-03-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.08386v1/

## Key points

- Humans and AI systems working together in hybrid teams are becoming more common as AI systems advance.
- While AI systems offer benefits, they are not error-free, making hybrid teams a suitable solution for maximizing performance.
- A manager model is introduced to delegate decision-making responsibilities among agents (humans and AI systems) over time using Reinforcement Learning.
- The manager's learning process is guided to minimize unnecessary changes in delegation due to undesirable team behavior.
- Experiments conducted in grid environments with failure states show the effectiveness of the manager in making optimal delegation decisions.
- The research explores related works such as Hybrid Reinforcement Learning and Delegation models that inform the approach.
- The model leverages concepts from multiple concurrently operating agents scenarios rather than following traditional hierarchical layers of abstraction seen in HRL approaches.
- Parallels are drawn with delegation models focusing on identifying suitable agents for tasks based on performance and cost considerations.
- Effective delegation strategies are highlighted as crucial for enhancing team performance within hybrid human-AI teams.

## Summary

In this paper, we delve into the concept of humans and AI systems working together in what is known as a hybrid team. As AI systems become more proficient and their adoption becomes widespread, the collaboration between humans and AI is expected to increase. However, it is important to note that while AI systems offer many benefits, they are not error-free. Therefore, hybrid teams are a suitable solution for maximizing performance. To enhance the efficiency of these teams, we introduce a manager model that learns how to delegate decision-making responsibilities among the agents (both humans and AI systems) over time. Our manager model utilizes a standard Reinforcement Learning scheme to determine the best agent to take on a particular task at any given moment. We also guide the manager's learning process to minimize unnecessary changes in delegation due to undesirable team behavior. Through experiments conducted in various grid environments with failure states, we demonstrate the effectiveness of our manager in making optimal delegation decisions. Furthermore, we explore related works such as Hybrid Reinforcement Learning and Delegation models that inform our approach. While our model does not follow the traditional hierarchical layers of abstraction seen in HRL approaches, it leverages concepts from multiple concurrently operating agents scenarios. Additionally, we draw parallels with delegation models that focus on identifying suitable agents for tasks based on performance and cost considerations. Overall, our research highlights the importance of effective delegation strategies in enhancing team performance within hybrid human-AI teams. By implementing our manager model  we showcase its ability to make informed delegation decisions that lead to optimal team paths while considering risk-based constraints. Our results demonstrate the efficacy of our approach in achieving desirable outcomes for hybrid teams operating in dynamic environments.