The ultimate goal of the roadmap exercise is to identify the research questions that need to be answered if we are to be able to build well-engineered cognitive systems and to identify them at quite fine levels of detail. This is challenging for many reasons: the people working in the discipline of cognitive systems are a very diverse group and this diversity is reflected in the scientific approaches they adopt as the basis for their research. It is crucial that a roadmap must not alienate sections of the community but it is equally crucial that the roadmap should be focused: the community doesn’t have infinite resources (funding, effort, and time) and we need to channel energy into the areas that show the greatest promise. The difficulty lies in identifying what these areas are, but to do so in a way that avoids polarization and that doesn’t further entrench existing positions. We need instead to highlight possible linkages: mutually relevant concerns that may allow the integration of ideas.

Cognitive systems display the common characteristic that they interact, they adapt, and they anticipate. However, every cognitive system is different, if only because cognition is effectively a means to an end: a way of overcoming the intrinsic uncertainty that accompanies interaction with a world that is not completely pre-specified whilst maintaining the integrity of the system (e.g. its autonomy) and still succeeding in achieving the systems purpose. It seems that it is these two issues that cause the greatest problem when it comes to building a roadmap: the purpose-specific (or goal/context/application/agent-specific) nature of cognition and the degree to which we would like the system to be independent of pre-specified knowledge. Depending on where you set the desired mark in these two dimensions, you will allow a certain type of cognition and, more importantly, a certain variety of possible underlying approaches.

The approach we had decided to adopt to address this challenge was to focus on the requirements of cognitive systems and then do some backward chaining to identify the gaps in our scientific and engineering knowledge which, if filled, would allows us to progress from where we are now to the point where we can build robust cognitive systems. This approach contrasts with the (normal) approach of focusing on our current capabilities and seeking to extend them, often quite incrementally. While there is clearly great merit in the backward chaining approach, the concern is that it can become an extremely large task, for example, by surveying the requirements of all possible applications areas and backward chaining to the complete set of underlying scientific principles. Indeed, we would need to answer almost all the questions posed in the Challenge 2 Background Document.

In addition to all this, there is also the the lack of maturity in the field to contend with. It is not only that we don't know how exactly what steps are required to link our current capabilities with our required cognitive capabilities (the identification of these is the purpose of the roadmap) but it is also the case that these steps have to be identifed in a scientific environment that is inherently immature and uncertain. At present, there is a lack interdisciplinary cohesion, there is a lack of a conceptual framework, and there is a lack of a shared application focus.