Improved AI Method Enables Reliable Logical Conclusions

Weighing up arguments, drawing logical conclusions and deriving a clearly correct answer – such tasks have so far presented artificial intelligence with a number of hurdles. When it comes to complex problems, computing a logically sound conclusion quickly pushes standard algorithms to their mathematical and computational limits. A team including Johannes P. Wallner from the Institute of Software Engineering and Artificial Intelligence at Graz University of Technology (TU Graz) has now found a way to translate the complex process of argumentative reasoning into more efficient mathematical formulas and to link these to established problem-solving systems. This enables precise calculations to be made in milliseconds. The team successfully tested the reliability of the method in collaboration with the Dutch police using a fraud intake tool.

Different from large language models

“Compared to large language models such as ChatGPT, which generate texts based on statistical probabilities and large amounts of training data, exact logical conclusions can quickly reach the limits of computability from a mathematical point of view,” says Johannes P. Wallner. “Here, the result must necessarily be based on the data specified for the problem in question, and the system must be guaranteed to provide the correct or best answer within the defined framework. Large amounts of data are a problem here, as the computational effort increases exponentially.”

To overcome this hurdle, the team focused on the mathematical modelling of the logical questions and developed precise logical formulas with which the AI system can independently find the optimal solution.

Online fraud as a use case

As part of the cooperation with the Dutch police, it was necessary to translate the problems involved in recognising potential fraud into compact formulas. The researchers had to provide mathematical proof that the formulae correctly reflected the reasoning for or against fraud in the data entered. The aim was to obtain an answer from the computer based on the available data and evidence as to whether a case could be internet fraud or not. This involved, for example, people who had bought a product via a website and received nothing in return. Considering the time elapsed since the order was placed, the webshop URL and other information, the system had to assess the case.

Fast enough for practical applications

Although this is primarily basic research, the researchers have identified numerous other possible applications in which complex decisions are made on the basis of logical arguments. A current challenge is still data input, as this has to be very precise for the system to be able to process it. In future, Wallner’s team will investigate connections with large language models. They could receive input in natural language and automatically translate it into a form suitable for logical processing.