AI Technion Lab
Between January 2022 and July 2022, I was employed as a student at the AI Technion Laboratory, working 2 to 3 days a week. During this period, I had the unique opportunity to work full-time on a Multi-Agent Reinforcement Learning project under the guidance of Dr. Davide Schaumann. This collaboration allowed me to explore an exciting new domain within reinforcement learning, specifically in the innovative subfield of multi-agent systems, where multiple AI agents learn to collaborate and function together. Over the course of six months, we made significant progress, including developing proofs of concept and building a simulator, which laid the groundwork for further advancements in this field.
The AI Technion Lab is an organization within the Technion that connects researchers from various fields with AI professionals, facilitating collaboration and the integration of AI tools into research across different laboratories. I had the opportunity to work with Dr. Davide Schaumann, a researcher from the Faculty of Architecture, who conducts research on optimizing building and resource management. Dr. Schaumann has developed several companies in this domain. The project he assigned to me involved implementing a new resource optimization model for a hospital department, aimed at maximizing patient utility, reducing treatment times, and optimizing the use of resources such as personnel, space, and equipment. This study served as a proof of concept to evaluate whether multi-agent reinforcement learning is a relevant approach for optimizing hospital resources.
The project revolves around the challenge of overcrowding and the resulting pressure on hospital staff. Dr. Davide Schaumann's hypothesis is that an AI-based solution can be developed to create an intelligent system that provides real-time recommendations to optimize the usage of resources within a hospital. This includes assigning specific tasks to staff, managing the use of equipment, and efficiently utilizing space in order to improve patient satisfaction, reduce waiting times, and effectively manage a variety of scenarios. In collaboration with Rambam Hospital in Haifa, a wealth of data has been provided by the catheterization department.
​
In a simplified context, focusing on a single type of surgery, Dr. Schaumann and a previous team had already achieved performance improvements using an A algorithm* to maximize patient utility in specific simulations. However, the computational time of a Greedy algorithm like A* yields optimal results but is too slow to provide real-time recommendations. The goal of this project is to explore a different approach, utilizing reinforcement learning algorithms to create a model that can approach the optimal results achieved by the previous team but with near-instantaneous computation time.
​
The project is divided into three key parts. First, a software engineering phase to develop a simulator using the OpenAI Gym environment. Second, the creation of a centralized reinforcement learning model that learns how to assign tasks while utilizing available resources efficiently. Lastly, the development of a multi-agent decentralized reinforcement learning algorithm, where each agent learns to coordinate with others to share hospital resources, ultimately aiming to maximize patient utility.