Scientists teach LLMS to deal with comprehensive challenges in the field of planning

The company seeks to minimize costs and at the same time meet 23 increase in demand.

Wouldn’t it be easy for a company to ask Chatgpt to come up with an optimal plan? In fact, for all their incredible abilities, large language models (LLM) often work badly when they are commissioned by a direct solution to sculptures of complicated problems with planning themselves.

Rather than trying to change the model so that LLM becomes a better planner, scientists have attracted a different approach. They introduced a framework that leads LLM, which breaks the problem as a human, and then automatically solves it using a powerful software tool.

The user must only describe the problem in examples specific to natural language — is required for training or challenge LLM. The text of the EN EN user into a format that can be intensified by the optimization solver designed to effectively break extremely heavy planning challenges.

During the LLM wording process, it checks its work on multiple intermediate steps to make sure the plan is correctly described for the solver. If it is a mistake rather than give up, LLM is trying to fix the broken part of the wording.

When scientists have tested their framework of nine complex challenges, such as minimizing remote stock robots, they must travel to complete tasks, have achieved a success of 85 Piernt, while the best base value has only achieved 39 pitch.

The versatile framework can be used on a number of multi -stage planning tasks such as planning crews of airlines or managing machine time in the factory.

“Our research introduces a framework that basically acts as intelligent persecution of planning problems. It can find out the best plan that meets all the needs you have, even if the rules are complicated or unusual,” says Yilun Hao, postgraduate student in the MIT laboratory for information and decision systems (Lid).

She joined her on Yang Zhang newspaper, science in the MIT-IBM Watson AI laboratory; and the head of Chuch Fan, associate professor of astronautics and astronautics and chief investigator. The research will be presented at an international conference on learning representations.

Optimization 101

The fan group develops algorithms that automatically solve what is called combinatorial optimization problems. These huge problematic parties can interconnected decision variables, each with more options that quickly add up to billions of potential options.

People will solve such problems by delaying several options and then determining which one leads to the best overall plan. Algorithmic researchers of scientists use the same principles for optimization problems that are too complex to crack.

However, the solvers they develop have a tendency to have steep learning curves and are usually used only by experts.

“We thought LLMS could allow notxperts to use these algorithms. In our laboratory, we take a domain expert and formalize it into the problem our solver can solve. We could teach LLM to do the same?” Says a fan.

Using the framework developed by scientists, called Formalized Programming (LLMFP) based on LLM (LLMFP), provides a description of the natural language of the problem, basic information about the task and the question that describes their goal.

LLMFP then invites LLM to consider the problem and determine the decision -making variables and key restrictions that will shape the optimal solution.

LLMFP asks LLM to describe in detail the requirements of each variable before coding information to the mathematical formulation of the problem of optimization. He writes a code that encodes the problem and calls the connected optimization solver that comes to the ideal solution.

“It is similar to teaching undergrades about problems with optimization at MIT. We do not teach them just one domain. We teach a thyodology,” Fan adds.

If the entrances to the solver are correct, it will provide the correct answer. Any errors in the solution come from errors in the wording process.

To ensure that he has found a work plan, LLMFP analyzes the solution and modifies any incorrect steps in the formulation of the problem. UNCE Plan This self -evaluation, the solution is described by the user in natural language.

Improving the plan

This self -esteem module also allows LLM to add any implicit restriction that missed for the first time, says Hao.

For example, if the framework optimizes the supplier chain to minimize coffee costs, one can now not make coffee negative amnt roasted beans, but LLM may not realize it.

A step of self -evaluation would indicate this error and call on the model to fix it.

“In addition, LLM can adapt to the user’s preferences. If the model realizes a specific use, it does not want to change the time or budget of its travel plans, it can propose a change!

In a number of tests, their framework has achieved an average success rate between 83 and 87 percent in nine different problems with the user planning of several LLM. While some basic models are better in sedimal problems, LLMFP has achieved the rate of Oveall ABIC as high as the basic techniques.

Unlike these other LLMFP approaches, it does not require examples of domain safety for training. It can find the optimal solution to the planning problem right out of the box.

In addition, the user can customize LLMFP for various optimization solvers by adding Fed instructions to LLM.

“With LLMS, we have the opportunity to create an interface that allows people to use Toms from other domains to solve problems in a way that they may not have thought about,” says Fan.

In the future, scientists want to enable LLMFP to take pictures as input to complement the descriptions of the planning problem. This would help to solve tasks that are particularly difficult to describe in a natural language.

This work was partly financed by the MIT-IBM Watson AI maritime research and laboratory.