AI systems are being trained across various fields to make meaningful decisions, such as using AI to manage city traffic to enhance speed, safety, and sustainability. However, this is challenging because reinforcement learning models often struggle with variations in tasks. To address this, MIT researchers have developed a more efficient algorithm to train these models. This algorithm strategically selects the most impactful tasks for training the AI, which maximizes performance and minimizes costs. For example, in controlling city traffic signals, it focuses on a smaller number of key intersections to train on, improving overall effectiveness. The researchers found this method to be 5 to 50 times more efficient than conventional approaches, leading to faster learning and better AI performance. Senior author Cathy Wu highlights the simplicity and effectiveness of their algorithm, emphasizing its potential for wider adoption.

The research, presented at the Conference on Neural Information Processing Systems, was conducted by Jung-Hoon Cho, Vindula Jayawardana, Sirui Li, and Cathy Wu. Traditional methods involve either training a separate algorithm for each intersection or one for all, both with drawbacks. The new method finds a balance by using transfer learning to apply a trained model to new tasks without additional training, focusing on tasks that enhance overall algorithm performance. The developed Model-Based Transfer Learning (MBTL) algorithm estimates the benefit of training new tasks by modeling individual task performance and generalization across different tasks, selecting tasks that offer the highest gains. This approach significantly increases training efficiency, using much less data to arrive at the same solutions. Tests showed MBTL's efficiency in various simulated tasks, achieving up to a 50x boost in training efficiency. This means greatly reduced data requirements for reaching optimal solutions. The researchers aim to expand MBTL to tackle more complex, real-world problems, particularly in next-generation mobility systems. The research received support from several institutions, including a National Science Foundation CAREER Award and an Amazon Robotics PhD Fellowship.