The Second International Workshop on Guided Self-Organisation (GSO-2009)

Abstracts

Paolo Arena  (Università degli studi di Catania, Italy)
Thursday, August 20, 2009
Spatial temporal patterns in nonlinear dynamical systems: from locomotion to perception in biorobots
The presentation will review an approach, based on nonlinear, locally connected dynamical systems, know as Cellular Nonlinear Networks (CNNs), for the emergence of dynamical oscillatory slow-fast patterns, useful for locomotion control in biologically inspired robots. Furthermore, the use of the same CNN structure will be shown to give rise to other kinds of patterns, representing equilibria in attractor based networks, useful for the design of a perceptual structure in biorobotics. The methodology will be presented together with some experimental results.

Nils Bertschinger  (MPI MiS Leipzig, Germany)
Tuesday, August 18, 2009
Autonomy and closure
Closure and Autonomy play a fundamental role in systems theory were they are used to define a system as distinct from its environment. From a more practical point of view, in robotic systems the question of how to obtain an agent with true behavioral autonomy is of great interest.
Here, an information theoretic formalization of informational closure and interactive autonomy is presented. Having well-defined, formal measures allows to quantify the degree of closure and autonomy that is found in natural as well as artificial systems. Furthermore, a close relation between the concepts of closure and autonomy is established.
Both concepts are illustrated using computer simulations of simple finite state automata. In the end, further applications especially addressing closure of the sensory-motor loop and self-organization of behavior are discussed.

Ralf Der  (MPI MiS Leipzig, Germany)
Tuesday, August 18, 2009
Self-organization of behavior for autonomous robots
The talk discusses how general principles of self-organization, like information maximization, can be realized in practice. We consider embodied robots with 10 - 30 independent degrees of freedom controlled by a neural network. Under a closed coupling paradigm, the behavior of the robot is defined by the several hundred parameters of that network. Behavior development is defined as a gradient flow on objective functions obtained from both information and dynamical systems theory. We describe techniques and methods how the gradients can be estimated in these extremely high parameter spaces in real time. Results are demonstrated by a number of robotic applications.

Frank Güttler  (Universität Leipzig, Germany)
Thursday, August 20, 2009
Predictive Information for self-organizing robots using wireless connected embedded-controller-boards: a technical point of view
The predictive information as a measure for the behaviour of self-organizing robots enables on the one hand the ability to evaluate their cognitive abilities. On the other hand the neural net controlling the robot is able to use this measure as an error function in order to obtain a good behaviour. The purpose to applicate the theoretical results to real robots is limited by several constraints such as computation power or time-critical implementation complexity. An introduction of a solution of a hardware platform for real robots shows how this problems are dealed successfully with a matter.

J. Michael Herrmann  (University of Edinburgh, United Kingdom)
Thursday, August 20, 2009
Guided self-organization for control of bipedal walking
While efficient walking can be achieved by the exploitation of passive dynamics, stability and flexibility of the gaits require adaptive control and remains a challenge in robotic and prosthetic applications. The complexity of the dynamics suggests a hybrid approach that we design by a genetic algorithm for the evolution of a system of low-dimensional control structures, a homeokinetic exploration of the controllable movements and a reinforcement algorithm for the goal-oriented development of the generated elementary behaviors. The results of the study are demonstrated by a physically realistic three dimensional model of a walking robot.

Joseph Lizier  (CSIRO ICT Centre, Australia)
Wednesday, August 19, 2009
Functional and structural topologies in evolved neural networks
The topic of evolutionary trends in complexity has drawn much controversy in the artificial life community. Rather than investigate the evolution of overall complexity, here we investigate the evolution of topology of networks in the Polyworld artificial life system. Our investigation encompasses both the actual structure of neural networks of agents in this system, and logical or functional networks inferred from statistical dependencies between nodes in the networks. We find interesting trends across several topological measures, which together imply a trend of more integrated activity across the networks (with the networks taking on a more "small-world" character) with evolutionary time.

Georg Martius  (MPI for Dynamics and Self-Organization, Germany)
Thursday, August 20, 2009
Goal-oriented behavior from guided self-organization of sensorimotor loops
We start from the self-organized development of behaviors based on the homeokinetic principle and demonstrate the emergence of playful and embodied behaviors using a variety of robotic systems. We present two ways to guide the self-organization process towards desired behaviors. The first method integrates an online reward signal into the framework to modulate the speed of search in the behavior space. This increases the likelihood to find and perform rewarded behavior. We validate this mechanism using a spherical robot. Second, we propose a mechanism to specify symmetries of the physical system or of the desired behaviors as soft-constraints. This reduces the effective dimensionality of the system and leads to an efficient self-exploration, which we illustrate using high-dimensional robots.

Oliver Obst  (CSIRO ICT Centre, Australia)
Tuesday, August 18, 2009
Information transfer in recurrent neural networks
Reservoir computing (RC) is a recent paradigm in the field of recurrent neural networks. RC computing approaches have been employed as mathematical models for generic neural microcircuits, to investigate and explain computations in neocortical columns. A key element of reservoir computing approaches is the randomly constructed, fixed hidden layer - typically, only connections to output units are trained. In previous work, we have addressed performance issues of Echo State Networks, a particular reservoir computing approach, and investigated methods to optimize for longer short-term memory capacity or prediction of highly non-linear mappings. A general method for improving network performance is the use of permutation matrices for reservoir connectivity, however, problem specific methods such as unsupervised learning based on intrinsic plasticity (IP) also exist. IP aims to increase the entropy of each output of the internal units, but unfortunately improves performance only slightly compared to a setup based on random or permutation matrices. Comparing completely random networks, and networks based on permutation matrices, we found transfer entropy between network input and output of individual units to be a significant indicator for the performance of the network. Higher transfer entropy seems to indicate a more homogeneous, coherent computation using permutation matrices as a result of the lower in-degree of nodes. A future extension of this work is to investigate methods to increase transfer entropies based on local learning rules in individual nodes.

Mahendra Piraveenan  (CSIRO ICT Centre, Australia)
Wednesday, August 19, 2009
Assortativity and information in directed biological networks
We analyse the relationship between assortative mixing and information content in biological networks which are typically directed. We develop the theoretical background for analysing mixing patterns in directed networks before applying them to specific biological networks. Two new quantities are introduced, namely the in-assortativity and the out-assortativity, which are shown to be useful in quantifying assortative mixing in directed networks. We also introduce a general measure for information content in directed networks, followed by the 'out-information' and 'in-information' to quantify the information content of out-degree and in-degree mixing patterns respectively. We apply the measures introduced to a range of real world networks, demonstrating that out-degree mixing patterns contain the highest amount of information in most real world biological networks.

Daniel Polani  (University of Hertfordshire, United Kingdom)
Tuesday, August 18, 2009
Be empowered: guiding adaptation through potential information flows
(joint work with Alexander Klyubin and Chrystopher Nehaniv)
The central importance of (Shannon) information as a resource for the adaptation of living organisms or agents has been increasingly established in the last years. The information perspective for the characterization of an agent's operation is highly attractive since it provides a universal currency and language for any kind of "information processing" taking place both within the agent and in the dynamics of its interaction with the environment, and allows one to adopt an extreme bottom-up view. Furthermore, it is "coordinate-free" in the sense that it allows to formulate principles and "balance sheets" without having to refer to a particular information processing mechanism
The use of Causal Bayesian Networks (CBNs) has been established as a successful technique to create informational models of agents and their perception-action loop. This technique allows the tracking of information flows through the composite agent-environment system, the generalization of Ashby's Law of Requisite Variety, or the application of generalized Infomax principles. In particular, the latter provide a path for the generation of structured information processing architectures with no assumptions beyond the agent being "embodied" in some structured environment. Phenomena such as active sensing emerge from the principle as a natural side effect.
The transparency of modelling the perception-action loop using the CBN formalism allows one to identify additional phenomena and quantities of interest. Specifically, in the present talk, I will introduce and discuss "empowerment" which is essentially the amount of potential information that an agent could inject into the environment via its actuators and recapture via its sensors. In the simplest of cases, this reduces to an agent-external channel capacity, but in general one requires CBNs to formulate empowerment.
In a situation where an agent has no prior preferences, its empowerment turns out to provide a utility which draws it to "interesting" states in the system. Since empowerment only depends on the embodiment of the agent, it can assign sensible preferences to states even in absence of any other prespecified drives (quantities of this kind we term "universal utility"). Understanding properties of possible universal utilities is particularly relevant for the success of adaptive systems, as the latter frequently have to be able to cope with novel situations that have not been previously encountered and for which the systems' innate drives are not appropriate or suitable drives may not exist yet at all.
I will show how, in a number of scenarios of varied quality and characteristics, the behaviour resulting from empowerment optimization is close to our intuitive expectations, sometimes achieved in a nontrivial way. In the discussion, I will suggest possible reasons for this and discuss lines for future research.

Mikhail Prokopenko  (CSIRO ICT Centre, Australia)
Thursday, August 20, 2009
Information transfer and cortical interactions in a visuomotor tracking task
The human brain undertakes highly sophisticated information processing facilitated by the interaction between its sub-regions. We present a method of identifying directed inter-regional information structure (using extensions to the transfer entropy), and the changes in this structure with respect to some variable (e.g. time). This method is distinguished in using asymmetric, multivariate, information-theoretical analysis, which captures not only non-linear relationships, but also collective interactions and the direction of these relationships. The method is used to analyse blood oxygen level-dependent time series to establish the directed information structure between regions involved in a visuomotor tracking task. Importantly, this is a tiered structure, with known movement planning regions driving visual and motor control regions. Also, we examine the changes in this structure as the difficulty of the tracking task is increased, and find greater coupling between regions involved in movement planning (left SMA and left PMd) and execution (right cerebellum for right hand and right SC for eye movements) with task difficulty. It is likely these methods will find utility in identifying inter-regional structure (and structural changes) in other cognitive tasks.

Mikhail Prokopenko  (CSIRO ICT Centre, Australia)
Wednesday, August 19, 2009
Information-theoretic modelling of scaling in genetic code
The principle of least effort in communications has been shown, by Ferrer i Cancho and Sol'e, to explain emergence of power laws (e.g., Zipf's law) in human languages. This study applies the principle and the information-theoretic model of Ferrer i Cancho and Sol´e to genetic coding. The application of the principle is achieved via equating the ambiguity of signals used by “speakers” with codon usage, on the one hand, and the effort of “hearers” with needs of amino acid translation mechanics, on the other hand. The re-interpreted model captures the case of the typical (vertical) gene transfer, and confirms that Zipf's law can be found in the transition between referentially useless systems (i.e., ambiguous genetic coding) and indexical reference systems (i.e., zero-redundancy genetic coding). As with linguistic symbols, arranging genetic codes according to Zipf's law is observed to be the optimal solution for maximising the referential power under the effort constraints. Thus, the model identifies the origins of scaling in genetic coding — via a trade-off between codon usage and needs of amino acid translation. Furthermore, the paper extends Ferrer i Cancho – Sol'e model to multiple inputs, reaching out toward the case of horizontal gene transfer (HGT) where multiple contributors may share the same genetic coding. Importantly, the extended model also leads to a sharp transition between referentially useless systems (ambiguous HGT) and indexical reference systems (zero-redundancy HGT). Zipf's law is also observed to be the optimal solution in the HGT case.

Susanne Still  (University of Hawaii, USA)
Tuesday, August 18, 2009
Interactive learning
I present a quantitative approach to interactive learning and adaptive behavior which integrates model- and decision-making into one theoretical framework. This approach follows simple principles by requiring that the observer’s behavior and the observer’s internal representation of the world should result in maximal predictive power at minimal complexity. Classes of optimal action policies and of optimal models can be derived from an objective function that reflects this trade-off between prediction and complexity. The resulting optimal models then summarize, at different levels of abstraction, the process’s causal organization in the presence of the learner’s actions. A fundamental consequence of the proposed principle is that the learner’s optimal action policies have the emerging property that they balance exploration and control. Interestingly, the explorative component is present in the absence of policy randomness, i.e. in the optimal deterministic behavior. This is a direct result of requiring maximal predictive power in the presence of feedback. Exploration is therefore not the same as policy randomization. This stands in contrast to, for example, Boltzmann exploration which is used in Reinforcement Learning (RL). Time permitting, I will discuss what happens when one includes rewards, such as is popular in RL.

Ivan Tanev  (Doshisha University, Japan)
Wednesday, August 19, 2009
Incremental genetic programming incorporating genetic transpositions for efficient coevolution of locomotion and sensing of simulated snake-like robot
Genetic transposition is a process of moving sequences of DNA to different positions within the genome of a single cell. It is recognized that the transposons (the jumping genes), facilitate the evolution of increasingly complex forms of life by providing the creative playground for the mutation where the latter could experiment with developing novel genetic structures without the risk of damaging the already existing, well functioning genome. In this work we investigate the effect of genetic transposition on the efficiency of genetic programming employed for coevolution of locomotion gaits and sensing of the simulated snake like robot (Snakebot). In the proposed approach, at the initial stage of evolution the pool of already evolved genotypes that control the locomotion of fast, yet sensorless Snakebots is first subjected to genetic transposition and then, the transposons are mutated in order to allow for the incorporation of the sensing information into the control of the bot. Experimental results suggest that the incremental grow of the genotype via genetic transposition and mutation, followed by the coevolution of the resulted locomotion control and sensing morphology contributes to the significant increase of the efficiency of evolution of fast moving Snakebots in challenging environments.

Christian Tetzlaff  (Max Planck Institute for Dynamics and Self-Organization, Germany)
Wednesday, August 19, 2009
Self-organized criticality of developing artificial neuronal networks and dissociated cell cultures
Self-organized criticality (SOC) [1] was first described in neuronal cell cultures by Beggs & Plenz [2]. Neuronal networks being in a critical state produce avalanche-like discharges that are power-law distributed. The assessment of avalanches in neuronal networks is a new way of looking at neuronal activities apart from bursts, synchronization etc. The main novelty of our approach is to assess the avalanche distribution at different developmental stages of neuronal networks. For this, we used dissociated post-natal cell culture taken from the rat cortex (Experimental data was provided by the Ulrich Egert group, BCCN Fribourg, Germany). We found that different network states as subcritical, critical or supracritical specify a time and spatial activity profile that is linked but not equivalent to low, moderate or high levels in neuronal activity, respectively. We are the first who show that the activity profile in cell cultures develop from supracritical states over subcritical into critical states. To shed light to the dependency of SOC on network development, we used a self-organizing artificial neuronal network model based on a previous model by Van Ooyen and Abbott [3, 4, 5]. An important novelty of our model is that it is more detailed with respect to representing seperate axonal and dendritic fields [6, 7]. The model network aims to develop towards a homeostatic equilibrium in neuronal activity which is achieved by growth and retraction of axonal and dendritic fields. This abstract model already reproduces the transient behaviour as seen in cell cultures from supracritical over subcritical to critical states. However, we found that some cell cultures remain in a subcritical regime. The model offers a simple explanation as depending on the strength of inhibition, equivalent to the friction in self-organizing systems [8], neuronal networks may or may not reach criticality even though they are homeostatically equilibrated.

References
1. Bak P, Tang C, Wiesenfeld K: Self-organized criticality: An explanation of 1/f-noise. Phys Rev Lett 1987, 59:381-384.
2. Beggs J, Plenz D: Neuronal avalanches in neocortical circuits. J Neurosci 2003, 23(35):11167-11177.
3. Van Ooyen A, Van Pelt J: Activity-dependent outgrowth of neurons and overshoot phenomena in developing neural networks. J Theor Biol 1994, 167:27-43.
4. Van Ooyen A, Van Pelt J, Corner M: Implications of activity-dependent neurite outgrowth for neuronal morphology and network development. J Theor Biol 1995, 172:63-82.
5. Abbott L, Rohrkemper R: A single growth model constructs critical avalanche networks. Prog Brain Res 2007, 165:9-6
6. Butz M, Teuchert-Noodt G, Grafen K, Van Ooyen, A: Inverse relationship between adult hippocampal cell proliferation and synaptic rewiring in the dentate gyrus. Hippocampus 2008, 18(9):879-898.
7. Butz M, Wörgötter F, Van Ooyen A: Activity-dependent structural plasticity. Brain Res Rev 2009, doi:10.1016/j.brainresrev.2008.12.023.
8. Lauritsen K, Zapperi S, Stanley H: Self-organized branching process: Avalanche models with dissipation. Phys Rev E 1996, 54:2483-2488.

Florentin Wörgötter  (University of Göttingen, Germany)
Thursday, August 20, 2009
Self-organized adaptation of simple neural circuits enables complex robot behavior
The control of complex sensori-motor systems is a challenging combinatorial problem because multiple simultaneous sensory signals need to be appropriately coordinated to yield a broad spectrum of distinct behavioral patterns. In this talk I will present a novel strategy to adaptively generate complex behavior of an autonomous robot using chaos-control in only one simple two-neuron module. The robot is sensor-driven by 18 inputs, which via a control network target 18 motors, thereby generating eleven basic behavioral patterns (e.g., orienting, taxis, self-protection, various gaits) and their combinations. The control strategy is adaptive and freely configurable by synaptic learning and thus provides an efficient yet simple way to self-organize versatile behaviors in autonomous systems with many degrees of freedom

Larry Yaeger  (Indiana University, Bloomington, USA)
Wednesday, August 19, 2009
Determining function from structure in neural networks
Previous work has investigated evolutionary trends in an information-theoretic measure of neural complexity, and related those trends to behavioral adaptation to the environment. This ("TSE") complexity measure thus appears to quantify the dynamical function of neural networks in an evolutionarily meaningful way. Other work has demonstrated that trends in increasing complexity are accompanied by corresponding trends toward increased clustering coefficient and decreased average minimum path length in the graphs that describe the underlying neural architectures. These results suggest an evolutionary trend towards so-called "small world" networks, and a correlation between small-world-ness and complexity. After a review of these results, I will ask the question: What other graph-theoretic metrics can we examine to further illuminate the relationship between network structure and network function, and what might they tell us about biological brains?

Keyan Zahedi  (MPI MiS Leipzig, Germany)
Thursday, August 20, 2009
Maximizing information in the sensory-motor loop
Cognitive systems are embodied and situated, which means that they act and learn within the sensori-motor loop. An open question is how such systems can improve their behaviour through self-organised learning. In this talk, we present an approach in which such a learning rule is derived from the principle of maximising an approximation of the predictive information. First experiments already show that such a principle enables the co-ordinated behaviour of physically coupled robots.

Date and Location

August 18 - 20, 2009
Max Planck Institute for Mathematics in the Sciences
Inselstraße 22
04103 Leipzig
Germany
see travel instructions

Scientific Organizers

Nihat Ay
Max Planck Institute for Mathematics in the Sciences
Leipzig
Contact by Email

Ralf Der
Max Planck Institute for Mathematics in the Sciences
Leipzig
Contact by Email

Mikhail Prokopenko
CSIRO
Sydney
Contact by Email

Administrative Contact

Antje Vandenberg
Max Planck Institute for Mathematics in the Sciences
Contact by Email
Phone: (++49)-(0)341-9959-552
Fax: (++49)-(0)341-9959-555

16.03.2017, 15:39