Researchers develop an efficient way to train more reliable AI agents

Unfortunately, teaching an AI system to make good decisions is no easy task.

Reinforcement learning models, which underlie these AI decision-making systems, still often fail when faced with even small variations in the tasks they are trained to perform. In the case of traffic, a model might struggle to control a set of intersections with different speed limits, numbers of lanes, or traffic patterns.

To boost the reliability of reinforcement learning models for complex tasks with variability, MIT researchers have introduced a more efficient algorithm for training them.

The algorithm strategically selects the best tasks for training an AI agent so it can effectively perform all tasks in a collection of related tasks. In the case of traffic signal control, each task could be one intersection in a task space that includes all intersections in the city.

By focusing on a smaller number of intersections that contribute the most to the algorithm's overall effectiveness, this method maximizes performance while keeping the training cost low.

The researchers found that their technique was between five and 50 times more efficient than standard approaches on an array of simulated tasks. This gain in efficiency helps the algorithm learn a better solution in a faster manner, ultimately improving the performance of the AI agent.

"We were able to see incredible performance improvements, with a very simple algorithm, by thinking outside the box. An algorithm that is not very complicated stands a better chance of being adopted by the community because it is easier to implement and easier for others to understand," says senior author Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS), and a member of the Laboratory for Information and Decision Systems (LIDS).

She is joined on the paper by lead author Jung-Hoon Cho, a CEE graduate student; Vindula Jayawardana, a graduate student in the Department of Electrical Engineering and Computer Science (EECS); and Sirui Li, an IDSS graduate student. The research will be presented at the Conference on Neural Information Processing Systems.

Finding a middle ground

To train an algorithm to control traffic lights at many intersections in a city, an engineer would typically choose between two main approaches. She can train one algorithm for each intersection independently, using only that intersection's data, or train a larger algorithm using data from all intersections and then apply it to each one.

But each approach comes with its share of downsides. Training a separate algorithm for each task (such as a given intersection) is a time-consuming process that requires an enormous amount of data and computation, while training one algorithm for all tasks often leads to subpar performance.

Wu and her collaborators sought a sweet spot between these two approaches.

For their method, they choose a subset of tasks and train one algorithm for each task independently. Importantly, they strategically select individual tasks which are most likely to improve the algorithm's overall performance on all tasks.

They leverage a common trick from the reinforcement learning field called zero-shot transfer learning, in which an already trained model is applied to a new task without being further trained. With transfer learning, the model often performs remarkably well on the new neighbor task.

"We know it would be ideal to train on all the tasks, but we wondered if we could get away with training on a subset of those tasks, apply the result to all the tasks, and still see a performance increase," Wu says.

To identify which tasks they should select to maximize expected performance, the researchers developed an algorithm called Model-Based Transfer Learning (MBTL).

The MBTL algorithm has two pieces. For one, it models how well each algorithm would perform if it were trained independently on one task. Then it models how much each algorithm's performance would degrade if it were transferred to each other task, a concept known as generalization performance.

Explicitly modeling generalization performance allows MBTL to estimate the value of training on a new task.

MBTL does this sequentially, choosing the task which leads to the highest performance gain first, then selecting additional tasks that provide the biggest subsequent marginal improvements to overall performance.

Since MBTL only focuses on the most promising tasks, it can dramatically improve the efficiency of the training process.

Reducing training costs

When the researchers tested this technique on simulated tasks, including controlling traffic signals, managing real-time speed advisories, and executing several classic control tasks, it was five to 50 times more efficient than other methods.

This means they could arrive at the same solution by training on far less data. For instance, with a 50x efficiency boost, the MBTL algorithm could train on just two tasks and achieve the same performance as a standard method which uses data from 100 tasks.

"From the perspective of the two main approaches, that means data from the other 98 tasks was not necessary or that training on all 100 tasks is confusing to the algorithm, so the performance ends up worse than ours," Wu says.

With MBTL, adding even a small amount of additional training time could lead to much better performance.

In the future, the researchers plan to design MBTL algorithms that can extend to more complex problems, such as high-dimensional task spaces. They are also interested in applying their approach to real-world problems, especially in next-generation mobility systems.

The research is funded, in part, by a National Science Foundation CAREER Award, the Kwanjeong Educational Foundation PhD Scholarship Program, and an Amazon Robotics PhD Fellowship.