AI-Powered Trip Planning: MIT-IBM Watson AI Lab Unveils Personalized Travel Broker

Navigating the complexities of travel planning, from transportation and accommodations to meals and lodging, can be daunting. While large language models (LLMs) offer promise with their natural language interaction and information-gathering capabilities, they often falter with the logistical and mathematical reasoning required for multi-constrained problems like trip planning. A recent study found LLMs provide viable solutions only 4% of the time, even with additional tools and APIs.

Addressing this challenge, a research team from MIT and the MIT-IBM Watson AI Lab has developed a user-friendly framework designed to act as an AI travel broker. This innovative approach combines common LLMs with algorithms and a complete satisfiability solver to create realistic, logical, and comprehensive travel plans.

“We believe a lot of these planning problems are naturally a combinatorial optimization problem,” explains Chuchu Fan, associate professor in the MIT Department of Aeronautics and Astronautics (AeroAstro) and the Laboratory for Information and Decision Systems (LIDS), and a researcher in the MIT-IBM Watson AI Lab. Fan’s team focuses on applying machine learning, control theory, and formal methods to develop safe and verifiable control systems.

The key to this framework lies in its ability to translate a user’s natural language description of the problem into a format that a solver can handle. Solvers are mathematical tools that rigorously check if criteria can be met, but they typically require complex computer programming. By using LLMs as translators, the new technique allows users to plan trips in a timely manner without needing programming knowledge or travel research.

The “travel agent” operates in four key steps: The LLM (GPT-4, Claude-3, or Mistral-Large) parses the user’s prompt, noting preferences for budget, hotels, transportation, destinations, attractions, restaurants, and trip duration. These steps are converted into executable Python code, calling APIs like CitySearch and FlightSearch to collect data, and the SMT solver executes the steps. If a solution is found, the solver outputs the result to the LLM, which provides a coherent itinerary to the user.

If constraints cannot be met, the framework identifies conflicting constraints and suggests potential remedies to the user. The user can then modify their preferences until a valid plan is formulated.

The researchers tested their method against other baselines, including GPT-4 alone and a search algorithm optimizing for total cost. The new technique generally achieved over a 90% pass rate, compared to 10% or lower for the baselines. They also found that adding a JSON representation within the query step further improved the pass rates to 84.4-98.9%.

The MIT-IBM team also tested the framework on other domains, such as block picking, task allocation, the traveling salesman problem, and warehouse optimization, demonstrating its generalizability and robustness.

Co-authoring the paper with Fan are Yang Zhang of MIT-IBM Watson AI Lab, AeroAstro graduate student Yilun Hao, and graduate student Yongchao Chen of MIT LIDS and Harvard University. The work was presented at the Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics.