GPT assistants head out to the forest

More and more people are using AI-based assistants as a sounding board, both in their personal lives and at work. Now, the technology is also being tested in the forest. Within Mistra Digital Forest, language models that specialise in planning and decision-making in forestry are being developed.

In a short space of time, GPT-based language models have become an integral part of many people’s lives. A project within Mistra Digital Forest is currently exploring how to harness this technology, in order to develop GPT-based assistants that will specialise in planning and decision-making within the forestry sector.

– Initially, the aim is to test the potential of the technology as the basis of a decision-support tool for planners. Technically speaking, it is not a huge leap from the GPT assistants that many people use today. We use the same technology, but the difference lies in how tightly the model is controlled, we limit the information it’s allowed to use. The idea is to develop an assistant that can provide support out in the field, and act as a tool for interpreting the knowledge and the documentation that already exists within the sector, says Morgan Rossander, a researcher at Skogforsk.

Transparency, robustness and predictability are key characteristics 

At the heart of this solution lies Retrieval Augmented Generation (RAG), a well-established method in which the language model does not freely generate responses based on its general knowledge, but rather it retrieves answers from a restricted set of documents. In this case, this involves a specially constructed database containing relevant and quality-assured information.

Researchers are currently making adjustments behind the scenes in order to tackle the problems that many language models struggle with. These include ‘hallucinations’ – where the language model starts making up answers and, during longer conversations, gradually strays from its original task. The researchers’ solution to the problem is that with every new question the model is given the entire task again, with instructions on who it is, and what it is supposed to do.

- This makes the assistant significantly more robust and predictable. The model must also show where the information comes from by linking to the correct page in a PDF, for example. This ensures the transparency that is otherwise a challenge with language models. It is usually difficult to know how they arrived at their answer, says Morgan Rossander.

Demo version available 

There is considerable interest in this type of solution, and now forestry companies can start planning to test an initial demo version. Sara Nilsson, planning specialist at project partner Holmen:

– This project fits in well with our strategy of moving towards more data-driven decision support. A forest area planner makes a great many decisions in their work, and all the factors are not relevant to every area. Our hope for this type of support is that it will provide planners with relevant information regarding the factors that are important for a specific area. Our requirement is that the language model provides the same answer regardless of who asks, and that it does not come up with its own solutions.

One of the key challenges that remains to be addressed concerns how the wording of a question influences the answer. Broadly phrased questions often yield general answers, whereas industry-specific wording provides more relevant feedback. In the long term, this type of planning support should even be able to answer ‘general’ questions in a specific manner. One possible way forward is a multi-step solution where the model is allowed to reason, search for information in stages and gradually narrow it down to what is relevant, rather than attempting to solve everything in one single step.

Looking for ways to make the model accessible 

The researchers are now exploring various options for making the model available for wider testing in forest planning. One approach could be to share an operational model via a commercial platform such as OpenAI, for easy access. Another approach could be to share the documentation in a way that enables companies to implement a similar assistant, based on a language model of their choice.

- Development is moving so fast that, in this project alone, we are already on our third basic model. What was cutting-edge a year ago is now both cheaper and more stable. New opportunities for how the technology can be used in practice are constantly emerging, says Morgan Rossander.