CAMBRIDGE, MA, 22 NOVEMBER 2024 – MIT researchers have presented a new algorithm, model-based transfer learning (MBTL), which helps to train AI systems much faster for challenging tasks that involve several decisions.
The method has the potential to transform industries such as traffic management and robotics by making AI systems more cost-effective and efficient.
The MBTL paradigm allows reinforcement learning models to train on a limited number of tasks while being able to compute for a wider scope of problems. It is 5 to 50 times more efficient than conventional approaches.
Cathy Wu, Associate Professor at MIT and senior author of the study informs,
‘’we achieved amazing results while using a comparatively simple algorithm which is good for the adoption of the AI community.’’
Lead author Jung-Hoon Cho concluded:
“Decreased availability of data for training enables us to achieve higher efficiency without affecting Predics performance.’’
MBTL was found to require only 2 tasks for training and achieved the performance of the standard ways, which reports the use of 100 tasks.
This development has a wide variety of usages, such as smart traffic systems, where AI agents can operate city intersections to minimize traffic congestion and preserve the environment.
Researchers intend to broaden the scope of the algorithm’s application to include high-dimensional and real-life tasks.
The study, co-authored by Wu, Cho, and MIT graduate students Vindula Jayawardana and Sirui Li, will be presented at the Neural Information Processing Systems conference.
Source: https://news.mit.edu/2024/mit-researchers-develop-efficiency-training-more-reliable-ai-agents-1122
What is your reaction?
