User-Centric Logic-Based AI

An organisation’s greatest asset are often its employees, who posses a huge amount of knowledge and expertise about products and processes. We build AI systems that use this knowledge to help people make better and more efficient decisions.

This requires two key ingredients: readable knowledge representation formalisms and flexible forms of logical inference.

Readable knowledge representation formalisms

In “good old fashioned AI”, a knowledge engineer would interview domain experts and build a formal model of their knowledge to serve as the foundation for an AI system. However, communication between a knowledge engineer and a domain expert can be difficult: the lack of a common language causes misunderstandings, that may remain unidentified for a long time.

We propose to let domain experts and knowledge engineers build the formal model together, typically during a number of workshops at a whiteboard. Here, the domain experts should feel comfortable reading and writing the model. This avoids misunderstandings and allows the domain experts to take ownership of the model.

To make this possible, we avoid “scary” logical formulas with complicated syntax. Taking inspiration from the OMG group’s Decision Model and Notation (DMN) standard, we have developed cDMN: a “syntax-less”, table-based notation for complex domain knowledge, that — like the DMN standard itself — is explicitly intended to be usable by people without a background in IT or computer science.

Flexible interaction

When the user is faced with a difficult problem, the system should not just take over and solve the problem for her. After all, she might not trust the solution, or have some restrictions or preferences that the systems knows nothing about. Instead, we want to support the user to arrive at a suitable solution by herself. To this end, we provide a set of logical inference methods that can be applied in a flexible way to interactively assist the user. We combine these logical methods in such a way that the user is always in control of what happens and understands what is going on.

This approach is implemented in our interactive consultant system. This tools allows users to explore and interact with a knowledge-base. Moreover, it is completely generic: you can load it with knowledge about any domain and it will immediately help you with decisions in that domain.

Applications

Our goal is to build knowledge-based AI systems that work in practice. Therefore, we often collaborate with industrial partners to validate and refine our methods. In recent projects, we looked at applications in component design, product selection and configuration of financial products.

Try it out?

While our real-life applications can get quite complex, we can illustrate the essence of the approach by a small example. The board game set is played using cards (see the picture below) which have four properties: color, shape, fill and number. The goal is to find a set of three cards that have, for each of these properties, either all the same value, or all a different value. That is, three red cards is fine and so is a set of a green, a red and a purple card, but a set of two red cards with one green card is not ok. And the same holds for the other three properties.

The rules of this game can be written in our cDMN notation, using a number of simple tables such as:

These tables can then be loaded into our generic interactive consultant tool. In the case of the above configuration, three sets are possible. The interactive consultants tells you that all of these sets have all different numbers and all different shapes and that they all contain the card of one red empty diamond, and that none of them contain, e.g., the card with two red striped diamonds. The user can then further explore the possibilities, e.g., by selected a card that she wants to include in the set, or by choosing whether she wants a set with all the same colors or all different colors (both are possible). In this way, she can guide the tool to the particular kind of set she is interested in. Alternative, she can use the logical inference of model expansion to ask the tool to compute a set for her.

You can try out this interface yourself by clicking here.