RAISON DATA - Rule-Based AI Systems for Reasoning on Massive Data

Many real-life decision problems humans are confronted with require the integration of two types of thinking: fast sub-symbolic thinking where patterns and situations are quickly recognised, and slow, symbolic thinking where reasoning tasks are performed based on transferrable knowledge learned from others or derived. Machine-learning (ML) techniques, in particular artificial neural networks, have enjoyed tremendous progress and are used in a wide variety of applications. They are often classified as sub-symbolic AI and correspond to fast thinking. The same is true for data mining. For more complex decision tasks, it would be desirable to integrate ML with advanced symbolic reasoning techniques. Many symbolic reasoning techniques have been proposed. However, none are expressive enough and, at the same time, simple enough to be used for the encoding of human knowledge and to interact with rules obtained via ML. 

The overarching goal of this project is to significantly contribute to next-level AI by finding methods, and building systems, that perform hybrid problem-solving tasks automatically, based on symbolic expert knowledge combined with machine learning over massive amounts of available data. The specific goals of the project are: 

1. To design a conceptually simple rule-based reasoning formalism for use over big data that is expressive enough to incorporate a wealth of essential features such as ontological reasoning, probabilistic reasoning, database access, data integration, default reasoning, aggregate functions and others. The new rule language should allow highly efficient, possibly parallelizable deduction algorithms that work well in practice even with very large heterogeneous data collections.
2. To integrate the above formalism with ML methods so that machine-learned rules can interact with rules from human experts. For example, expert rules may complement or override ML rules, or guide the ML process.
3. To implement and test an efficient hybrid “bilateral” problem-solving system that integrates ML with transferable expert knowledge, and apply it to various real-life problems of corporate and institutional decision-making.

Progress towards these goals will mean progress towards a new level of more trustworthy AI that makes a more balanced use of both transferrable knowledge provided by humans and machine learning.