Joined-up reasoning for automated scientific discovery a position statement and research agenda

We use the phrase “joined-up ” here with a double mean-ing: to convey two aspects of scientific discovery which we believe are essential, yet under-researched with respect to automating scientific discovery processes. Firstly, from an Artificial Intelligence perspective, the majority of ap-proaches to using AI techniques involve a disjointed se-quential application of different problem solving methods, with the user providing the glue in various ways. These in-clude routine logistical aspects such as the pre-processing of data and knowledge, translating outputs into input, choos-ing parameter settings for running AI methods, etc. More importantly, however, the user performs various aspects of meta-level reasoning, including asking the most pertinent questions, determining what it means if a process terminates with success and identifying – and investigating – anoma-lies. This approach tends to lead to auto-assisted discov-eries where the user knows what they are looking for, but not what it looks like, rather than the deeper discoveries of examples/concepts/hypotheses/explanations that the user didn’t even know he or she was looking for. While AI meth-ods promise the discovery of such surprising and novel sci-entific artefacts, they rarely deliver on this promise, as their application is too regimented within the problem solving paradigm of AI. We therefore advocate (and actively pursue) investigations into how to build systems which combine reasoning activi-ties in such a way that the whole is more than a sum of the parts. While automating the logistical aspects mentioned above is tiresome but straightforward, we believe that au-tomating the meta-level reasoning skills employed by the users of AI tools for discovery tasks is a fascinating prob-lem. To this end, we have looked at various ad-hoc combina-tions for discovery tasks in pure mathematics (e.g., (Colton & Pease 2005), (Charnley, Colton, & Miguel 2006), with a summary in (Colton & Muggleton 2006)), but more recently we have started to investigate the value of more generic ap-proaches based on proof-planning from automated theorem proving (Sorge et al. 2007), and global workspace architec-