This category should contain all of the types of abstract reasoning that the end of CoSy system should be able to do that doesn't fit elsewhere. In particular this category will reference reasoning abilities concerned with the PlayMate's own behaviour.

One particular reasoning ability that must be present in the PlayMate by the end of the project is the ability to do counter-factual reasoning, i.e. to be able to reason about things that do not currently exists. There are a number of specific motivations for including this in the requirements. First, the Playmate must be able to reason about things it cannot see. This may be in the context of something obscured in its current view of the world, something potentially 'off-screen' in the current scene, or something that another agent may have removed from the scene. The PlayMate must also be able to reason about the effects its actions may have on the world, and in a related competence, produce plans for future actions. A related but possibly more advanced competence is the ability to explain events in the world by positing possible causes.

Although individual columns within this matrix will almost definitely have separate ontologies within the cosy system, it may be informative to consider the role of ontologies within the end-of-project system in general. To ensure that the various subsystems within the PlayMate are able to refer to the same abstract concepts, their interactions should be mediated by ontologies. This is an approach that was successfully demonstrated in the year 1 demo system. For the remainder of the project, it is an open question how we can extend this approach into more specialised sub-fields. For example, if the CoSy system must pick up an object, its manipulation system should use the same action ontology as any action recognition competence it has. And the object representation that it uses for manipulation should also be linked in some manner to the representation used for vision, language etc.

In an effort to investigate how meaning, experience and context can be used in an intelligent system, it is important that the cosy system has episodic memory. It is likely that elements (within episodes) within this memory will need to be linked (possibly via pointers) into the system's ontologies to associate the memories with known concepts.

All the above linking may seem like unnecessary, and messy, computational overhead, but by building such structures, we can investigate how an intelligent system can make use of the kind of densely interconnected knowledge that humans have. In particular this may be used to drive cross-architecture control mechanisms (such as visual attention), and goal generation.

In general, the problem of sharing information between processes is one that we must tackle in the short to medium term. The work in the first year highlighted the problems of keeping information privately encapsulated within processes and therefore removing shared information from its processing context. Moving beyond this is a challenge both in terms of architecture design, and in implementation.