Imagine a coffee company that is trying to optimize the supply chain. The source company of three suppliers, it will bake them in two objects in gloomy or delicate coffee, and then sends roasted coffee to three retail locations. Suppliers have different fixed capacity, and baking costs and shipping costs vary depending on the place.
The company tries to minimize costs when meeting a 23 % augment in demand.
Wouldn’t it be easier for the company to just ask Chatgpt for an optimal plan? In fact, despite all their amazing capabilities, huge language models (LLM) often work poorly when the task is to directly solve such complicated problems with planning yourself.
Instead of trying to change the model to make LLM a better planner, myth researchers adopted a different approach. They introduced the frames that lead LLM to break up the problem, just like man, and then automatically solve it with a powerful software tool.
The user only needs to describe the problem in natural language-they are not needed for the task to train or install LLM. The model codes the user’s text in the format, which can be solved by the optimization Solver designed to effectively break up extremely arduous challenges related to planning.
During the formulation process, LLM checks its work on many indirect steps to make sure that the plan is correctly described to Solver. If he notices a mistake and not surrender, LLM tries to fix the broken part of the preparation.
When scientists tested their frames on nine completed challenges, such as minimizing warehouse robots at a distance, they must travel to perform tasks, achieved an 85 -percentage success rate, while the best base line reached only a 39 -month success rate.
The versatile framework can be used to a number of multi -stage planning tasks, such as planning air crews or machine time management in the factory.
“Our research shows frames that generally acts as an intelligent assistant to planning problems. This can determine the best plan that meets all needs, even if the rules are complicated or unusual,” says Yilun Hao, a graduate of the MIT laboratory in the field of information and decision systems (LIDS) and the main author of the author A Article about these studies.
It is joined by the article by Yang Zhang, a scientist from MIT-IBM Watson Ai Lab; and the elderly author of the Chuch, associate professor of aeronautics and astronautics and the main researcher. The research will be presented at an international conference on the representation of learning.
Optimization 101
A group of fans develops algorithms that automatically solve the so -called combinatoric optimization problems. These huge problems have many related decision -making variables, each with many options that quickly sum up billions of potential choices.
People solve such problems, narrowing them to several options, and then specifying which leads to the best general plan. Solvery algorithmic scientists apply the same principles to optimization problems that are too complicated for human ones to crack.
But the solvers they develop usually have steep learning curves and are usually used only by experts.
“We thought that LLM may allow us not to use these solving algorithms. In our laboratory we accept the problem of a domain expert and formalize it in the problem that our Solver can solve. Can we teach LLM to do the same?” The fan says.
Using RAM developed by scientists, called formal programming based on LLM (LLMFP), a person presents a description of the problem, information about the task and the question describing their goal.
Then LLMFP encourages LLM to justify the problem and determine the decision -making variables and key restrictions that will shape the optimal solution.
LLMFP asks LLM to specify the requirements of each variable before coding information in the mathematical formulation of the optimization problem. He writes a code that codes for the problem and causes the attached SOLVE optimization, which comes to a perfect solution.
“This is similar to how we teach students about problems with optimization in myth. We do not teach them only one domain. We teach their methodology” – adds the fan.
As long as the input data to Solver is correct, it will give the right answer. All errors in the solution come from errors in the formulation process.
To make sure he found a work plan, LLMFP analyzes the solution and modifies any incorrect steps in formulating the problem. After passing this self -esteem, the solution is described to the user in natural language.
Plan improvement
Hao says that this self -esteem module also allows LLM to add all the closed restrictions that he overlooked for the first time.
For example, if the frame is optimized by the supply chain to minimize coffee costs, man knows that coffee cannot send a negative amount of roasted beans, but LLM may not happen.
A step of self -esteem would mean this error and will make the model to repair it.
“In addition, LLM can adapt to the user’s preferences. If the model realizes that a specific user does not like to change the time or budget of his travel plans, it may suggest changing things that match the needs of the user,” says the fan.
Unlike these other approaches, LLMFP does not require training examples specific to the domain. It can find the optimal solution to the planning problem immediately after removing from the box.
In addition, the user can adapt LLMFP to various optimization solutions, adapting hints powered to LLM.
“In the case of LLM, we have the opportunity to create an interface that allows people to use tools from other domains to solve problems in a way that they could not think about before,” says the fan.
In the future, scientists want to enable LLMFP to take pictures as a contribution to supplement the descriptions of the planning problem. This would assist the frame in solving tasks that are particularly arduous to fully describe in natural language.
These works were partly financed by the Office of Naval Research and MIT-IBM Watson Ai Lab.