What are the organization principles for systems using networks of specialized subunits for a wide spectrum of possible tasks? How do these principles guide the forming of functional structures ranging from cortical neural circuits to concepts of skill webs containing complex body movements and social interaction capabilities?
Evolution has produced cognitive systems that are capable of solving the most complex tasks while at the same time maintaining homeostasis and interacting successfully with their environment on different levels. Part of this functionality is hardwired into the systems’ architectures, and part of it is embodied in their ability to respond to environmental changes that allows for reorganisation and learning. Within these systems, specialised hierarchical assemblies of subsystems have been established, and connections between these subsystems link functional components within and across levels of hierarchy. The performance of the whole system emerges from the functionality of its subsystems, their connectivity, and interaction dynamics.
The research area Skill Webs covers a wide range of topics that deal with systems composed of networked components, each of which has a particular function within the system. Nodes of this network can be situated on different organizational levels, like macroscopic and microscopic biological systems: societies, specimens, organisms, cells, and molecules; or corresponding organizational levels in hierarchical artificial systems, such as multi-agent systems or knowledge-based systems. The scientific target of Skill Webs is to understand the organisational principles of such networks on the systems level by means of building computational models.
In Skill Webs, our research is motivated by questions like: How are the subsystems within biological cognitive systems interconnected? How do they interact? How – and why – have their specialised architectures been evolved, and what are the underlying evolutionary driving forces? What happens if these models are exposed to realistic or real environments in the presence of a large amount of uncertainties? By answering these questions, we hope to understand the relationship between structure, dynamics and functional behaviours, as well as the optimisation principles of biological or artificial cognitive systems. Research topics we want to address in Skill Webs are exemplified in the following.
Our focus in this field of research will be on analysing complex nervous systems such as human brains and understanding the way in which their parts are interconnected. In the human brain, neuronal pathways belonging to various functional systems connect cortical areas with each other and with subcortical areas, forming a multitude of interacting circuits. Cortical and cerebellar microcircuits act as specialised functional units on the neuronal level. Synaptic and functional connections are formed and broken up again, depending on the activity of the neuron and its interaction with other cells. Constant rewiring and functional plasticity are major characteristics of biological nervous systems and necessary prerequisites for learning on the systems level as well as for functional stability. In addition to understanding the organisational and functional principles of these biological systems, we are also interested in how we can learn from these systems and their intrinsic architectures for building robust artificial systems.
How different parts of biological systems, like morphological characteristics and corresponding control structures such as nervous systems, co-evolve under given environmental conditions is a question of particular interest. Successful movement requires not only suitable body parts for walking, crawling or swimming, but also the means for controlling their respective movement and coordinating their mutual influences. By simulating evolutionary processes in artificial systems, we want to investigate how subsystems are linked together in evolution and what mechanisms lead to optimized architectures. Another aspect we want to investigate is the mutual influence of learning, development, and evolutionary processes in changing environments, and the question how structural and control units in gene regulatory networks interact with each other in evolution and development.
Complex body movements (such as turning a somersault or dancing) and manual tasks (such as binding a knot or playing the piano) can be subdivided into partial actions that are embedded into a functional and temporal structure. In order to perform such a task, an agent has to accomplish the subtasks and connect and coordinate them with appropriate timing. Order formation on the cognitive level is achieved through training and reflects the quality of performance and thereby the level of expertise. One approach we want to take is to analyse the biomechanics of complex whole-body movements and manual tasks and investigate the corresponding cognitive control structures. Using sophisticated robotic hands, we want to simulate how grasping and object manipulation is influenced by knowledge about procedures, objects and their affordances. Focusing on the acquisition of tool-use, we want to ask how (novel) objects can be integrated into task-specific body models to facilitate the performance of new skills.
In human and animal societies, connections are made between individuals and groups, allowing for dynamic interaction on different time scales. To be able to act and react successfully in a realistic social environment, an artificial agent has to be able to learn from and about the behaviour of its co-agents. Analysing the structures of hybrid societies in which humans and robots coexist as fellow beings and comparing them to the structures of human and other primate societies enables us to improve the interaction skills of our robots to make them fully acceptable partners for various tasks. For smooth and successful interaction in real-time, interfaces are required that provide efficient and robust connections between human and artificial agents. When cooperating or competing with others, an agent has to take into account the actions of its co-agents and their movement in space, and ideally understand their intentions to anticipate future actions. In an additional approach, we want to apply artificial life scenarios to evolve, manipulate, challenge and evaluate skills and attributes that enable agents to interact, cooperate or compete successfully with each other.