Criticality and Self-Organization in Behavioral Control
The biological function of self-organized criticality is much less understood than the physical mechanisms behind this phenomenon. A stereotyped response to external stimuli would be less advantageous than a flexible reaction which may amplify barely noticeable events in the environment based on information which has been accumulated in the internal state of the brain. Critical dynamics seems thus beneficial to living beings and it is known to bring about optimal computational capabilities, optimal transmission and storage of information, and sensitivity to sensory stimuli. It could represent a state where many options are available to the organisms and which is effective in development, fall-back behaviors, and exploration. We are interested in the developmental aspects of motor behavior in animals which we study in biomorphic autonomous robots, see picture below. Here the results on the neural systems have to be extended to incorporate the interaction with the environment. In simplified examples it can be shown explicitly that criticality is achieved from an objective function that characterizes both the sensitivity of the behavior with respect to sensory inputs and the predictability of the sensory effect while performing a behavior. The question whether such principles are also effective in simple organisms is a topic of the cooperation in the MPI-DS.
Physically realistic simulation of a four-legged robot. From the objective of simultaneous sensitivity of the behavior with respect to sensory inputs and predictability of sensory effects of the behavior, the quadruped generates autonomously variety of behaviors that include rising and lowering, coordination of leg movements, and basic locomotory movements.
In physically realistic simulations of autonomous robots with various sensory configurations and body shapes we obtain a continuous flow of behaviors that explores the variety of interactions of the robot with its environment. There remains, however, a restriction to self-organized solutions of relatively simple control problems such as the coordination of limb movements. In order to achieve complex and goal-directed behaviors we consider an architecture where the low-level criticalization is complemented by a additional learning mechanisms which evaluate the learning progress of the low-level adaptation in various environmental situations [Hesse et al 2007] or which use external reward signals to select behaviors which are generated by the self-organizing controller [Martius et al 2007]. Presently, the controller is replaced by a multi-agent control system where individual agents engage in the control of the robot if their prediction of the current behavior of the robot is sufficiently reliable, see the following images:
Physically realistic simulation of a spherical robot which is driven by three internally movable masses.
When being controlled by a multi-agent system, the spherical robot (image above) develops representations of a number of natural behaviors where each behavior corresponds to one controlling agent. The thus emerging elementary behaviors are characterized by one fixed axis and a coordinated movement of the masses along the other two axes.
This approach revealed to us the similarity of pattern formation algorithms that were studied in the visual system [Mayer et al 2007] but seem to be relevant also for the development of motor representations, where the competition among agents leads to a specialization of the corresponding behaviors which can be interpreted as elementary behaviors. In order to connect the results of this project more closely to results from biological experiment as well as to evaluate the behaviors emerging in the robots, we are studying algorithms for the analysis of complex data, which focus on intrinsic significance measures for the detection of relevant data features [Herrmann et al 2007, Voultsidou et al 2007]. The interaction of the controller with the sensory system and memory is studied in an integrated architecture that was originally developed for attention mechanisms [Schorbsdorff et al 2007], but proved to be generalizable to control of bio-robots.
Members working within this Project:
Former Members:
Katja FiedlerFrank Hesse
Georg Martius
