Description of the project

The SOKS project attempts to combine two different (and until now separated) main strains in Artificial Intelligence research: computational intelligence and knowledge representation. Scientifically it amounts to applying bottom-up computational intelligence techniques to solve the problem of creating shared ontologies through emergence and self-organisation.

Until now, the knowledge representation and computational intelligence research communities have lived almost entirely separate lives. Separate conferences, separate journals, separate methodologies, separate deployment areas. Even inside the AI Dept. at the VU this is apparent, with strong representatives of each area (KRR group of Van Harmelen, CI group of Eiben), but with zero interaction between them. However, recently there is an increasing awareness that principles from self-organising systems must be brought to bear on KR systems. Some researchers have suggested that a self-organising approach to ontology-alignment would be fruitful. The applicants agree to this vision and see a high potential in research along this line. They are willing to take up the scientific challenge and build a joint expertise in a new area.

Roughly speaking the problem is to create a shared ontology for a possibly large group of independent, but interacting units, where

1. Units may share the underlying semantic framework, i.e., set of concepts, but use different labels for these concepts, or
2. Units may have different sets of concepts that can overlap, and additionally
3. Units may or may not be engaged in dynamically creating their own concepts, resp., ontologies

The essence of the research approach proposed here is to initiate emergent processes among the units that will create those shared ontologies from bottom-up. The primary candidates to base such processes upon are techniques from evolutionary computing, self-organizing systems, machine learning, distributed computing, etc. The necessary expertise is largely present in the two groups. KR group: semantic web, negotiating meanings and routing semantic queries. CI group: evolutionary algorithms, self-organisation, machine learning.

• Overall research question: Given a set of basic ontologies, can we create through a bottom-up, emergent procedure a set of mappings between them that allows integrated reasoning with the combination of the basic ontologies.
  Operational questions that must be decided are what mappings are allowed between the basic ontologies (e.g. equivalence, subsumption, approximate, others?), and how to measure the quality of the obtained result.

• Computation cost: Is this procedure computationally feasible? How do the computational costs scale with the sizes of the basic ontologies, with number of the basic ontologies, with the connectivity in the network, etc.
  The appropriate performance parameters must be determined as part of the research. (Due to the exploratory nature of the project).

• Methodology: What is a sound experimental validation methodology? One option is to use real-life data, for examply by adopting the benchmarking data from the annual Ontology Alignment Challenge (2004, 2005, 2006), where both data and answer sets are publicly available.
  Alternatively, we can use random generators that produce representative problem instances, given some of the scaleable parameters mentioned above.

The following four steps represent a work programme that is, of course, subject to adjustments on-the-fly.

1. Survey existing work and identify key assumptions and parameters: Because the goal of this proposal is to build up a position in this area (rather than build on an existing one), we start with a survey of existing work in the area of emergent Knowledge Representations, for instance gossiping-style approaches. This should lead to a survey of the state-of-the-art, including better understanding of the problem area, the identification of key assumptions behind the existing work, the parameters that determine the design-space of these systems, and an inventory of the main challenges.
2. Repeat some of the earlier experiments: As part of building up our expertise in this area, we will want to do not only a paper survey of work by others, but also repeat some of the earlier experiments ourselves. This will no doubt help in identification of key assumptions and parameters, will result in a critical appraisal of the results obtained by others, and will yield inspiration for the next phase.
3. Extend with new experiments: We expect that after the first year, we will be in a position to formulate new models, to formulate new predictions based on these models, and to design new experiments to validate these predictions.
4. Generalise beyond experiments: In the final year, we will want to take a step back from the model-building and experimentation in years 2 and 3 (step 4), and to see if we can make claims about the generalisability of our results beyond the specific settings of our experiments. Also, we will want to consider the philosophical aspects related to this work.