Problem definition
The aim of this research is to define, develop, and exploit artificial systems with quasi cognitive capabilities that are complementary to (as opposed to mimicking) human cognition. These properties will bring the interaction between humans and artificial systems beyond the mere presentation of explicit input and output data, and will support effective collaboration at the level of problem formulation, analysis, and solution generation. Such a synergy between humans and artificial systems paves the road for defining effective principles for purposeful self-organization, where hybrid collectives of human experts (“actors”) and artificial systems with complementary cognitive capabilities (“cognitive agents”) naturally emerge, triggered by a certain problem, and reaching a structural and functional complexity that optimally matches the size and nature of the problem. These Actor-Agent Communities (AAC) are at the core of a new computing paradigm that is expected to overcome the well-known limitations of the current “Turing machine” paradigm (e.g. inadequacy for treating intractable, undecidable, and non-computable problems), summarized by Penrose (Shadows of the Mind, 1994) as: “[...] insights that are available to […] anyone who can think logically with understanding and imagination – lie beyond anything that can be formalized as a set of rules.”
The AAC concept proposed here overcomes the said limitations of the Turing paradigm not by changing the bases of algorithmic computation, but rather by a flexible and dynamic combination of human expertise and basic algorithms. We will illustrate an operationalization of this new concept in the practical application domain of crisis management, by gradually developing the key theoretical concepts and the technical solutions for the concrete requirements of these applications.
 

AAC Problem Solving
Actor-Agent Communities are (complex) information systems that can produce solutions for otherwise intractable information processing problems. They can be described as groups of humans and artificial systems that work together to achieve a common goal (i.e. solve a problem).
At the highest abstraction level an AAC can be regarded as a collection of autonomous problem-solving entities, with specific problem-solving capabilities resulting from their individual knowledge models, interaction modalities, and limited capabilities and resources for information processing.
A complex problem is one that cannot be decomposed into a finite number of simpler problems that would fit the solving capabilities of the available entities, either because of lacking adequate models, or because solving it requires general intelligence. For these reasons, a complex problem cannot be solved (yet) by machines.
The common approach to dealing with the first category of problems relies on approximation through heuristics. In data-driven approaches these heuristics can in principle be found through training and learning algorithms, provided that enough training data is available. However, the utility of these solutions is mostly limited to classification problems.
For those problems whose solving requires general intelligence, i.e. the capability of creatively combining pre-existing knowledge and data for solving a new (i.e. not previously encountered or described) problem, no automation is possible yet.
From this perspective, there is a need for at least two categories of entities: some that are capable of fast and reliable processing of large amounts of data, and others that benefit from general intelligence, hence the obvious choice for a hybrid system-of-systems.
 

AAC Self-Organization
A coherent behaviour can be induced on an amorphous collection of such autonomous entities by assigning them a common goal. This may be regarded as a set of quantities that describe a particular state of the environment (possibly including states of the entities themselves). Depending on the richness and complexity of their respective world models, each entity will maintain a particular (partial) representation of these goals. Problems can now be defined as mismatches between a target state of the environment and an observation thereof.
Such problems can be detected by one or more entities. However, the entity that discovered the problem may not be able to solve it all by itself, in which case other entities will be asked to contribute. A subset of entities forms thus a team whose purpose is solving that particular problem. In order for an entity to be able to contribute to solving the problems detected by other entities, its world model needs to partially overlap with those of the requesting entities. Indeed, this overlap is the obvious “common ground” required for any kind of collaboration.
This kind of self-organization is only constrained by a single property, that of belonging to the goal-driven community formed by the given goal. For the rest, the initiation and completion of self-organization is only determined by the detected problem in the environment and the capabilities of the contributing entities. In principle, such a team may dissolve after restoring the state of the environment. However, entities may be endowed with an adaptation mechanism that allows them to “remember” past collaborations, which would speed up team formation when similar problems occur in the future.