Trust your agents- mixed human-machine cooperation based on model-based, hierarchical and communication-augmented Reinforcement Learning
Abstract
Collaboration between humans and machines is a long-standing goal given the enormous advantages it can bring in terms of productivity increase, workload reduction, and general improvement of the quality of life for humans. Recently, there has been a widespread shift toward human-centric Artificial Intelligence (AI), where AI is seen as complementary to human abilities rather than as a substitute. The novel challenge is to show that the new paradigm of AI-empowered people is more effective than AI-only or human-only systems when it comes to completing a collaborative task. Classical approaches to modeling the machine side in the human-machine teaming usually rely on predefined and deterministic rules and are not scalable with respect to the number of state variables. Deep Reinforcement Learning (DRL) solutions on the other hand have shown their effectiveness in solving problems with huge state and action spaces and their ability to generalize to unseen test scenarios, although still limited to application scenarios where large amounts of data can be gathered. The research described in this proposal aims at studying novel models and algorithms for model-based and hierarchical multi-agent deep reinforcement learning, which is a completely new field of investigation. The hierarchical decomposition of a high complex scenario will allow the agents to reason on multiple levels of abstraction and to share only coarse-grained information with human collaborators, while still being able to share more fine-grained or low-level information with machine collaborators. The use of model-based techniques will allow an agent to build an internal representation of the environment and of the other agents in it (both artificial and humans). Using appropriate and interpretable communication systems, such as images, text or audio, the agents can transmit their internal representations to human teammates, who can thus acquire greater confidence in the abilities of their machine partner. Moreover, the prediction capabilities of the model-based approach combined with the explainable communication will allow humans to build trust in their artificial teammates. Finally, the learned models can help in explaining, monitoring and possibly driving the factors that most influence human behaviors.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 22, 2024
- Source ID
- FA86552317257
Entities
People
- Luca Iocchi
Organizations
- Air Force Office of Scientific Research
- Sapienza University of Rome
- United States Air Force