Cognitive systems are a new progressive area that appeared at the intersection of investigations in distributed artificial intelligence, machine learning and neurobiology. In this paper is briefly described a structure of a complex cognitive control system and suggested three possible models of the activator (image processing module in the cell).

Когнитивные системы – это новое прогрессивное направление, возникшее на стыке исследований в областях распределенного искусственного интеллекта, машинного обучения и нейробиологии. В докладе рассматривается обобщенная структура сложно организованной когнитивной управляющей системы, и предлагаются возможные модели активатора (внутриклеточного модуля процессирования образа).

Cognitive control systems (CCS) are a new class of the learning intelligence systems [1]. In general the functional network components of the CCS are realized on neural like elements which form structures with four levels:

That structure can formally be represented by the relation “cons” (connect of structure) [2]:

CCS_s = (CN₁,…, CN_i,…, CN_I) ;

CN_i = cons(…, PN_j^[1](…, CC_k^[2](…, CC_l^[3](…, CC_m^[4], …)…)…)…) ;

CC_m^[4] = C ; m=1,…,M ;

CC_l^[3] = C or N ; l=1,…,L ;

CC_k^[2] = C or N or LN ; k=1,…,K ;

PN_j^[1] = C and N and LN ; j=1,…,J , ( 1 )

where CCS_S – the structure of the cognitive control systems (CCS); CN_i (i=1,…,I) – overlapped networks; CC - cognitive center; C - cell; N - nucleus; LN - local network; PN - projective network.

The upper indexes designate the appropriate nested layer.

It can be offered three variants of CCS topology accordingly to a type of links between cells:

1. The systems with direct images transformation. Transformation of images occurs only in one direction - from inputs to outputs. Once upon a time learned, CCS of direct transformation is stable and can recognize images. There is no dynamic in such systems. Consequently, in such systems doesn’t appear a problem of stability. System with direct transformation is a static non-linear converter.

2. The systems with reconversion of images. It’s a dynamic system in which recurrent transformation is produced. Such system appears the problem of stability, which in biological analogues solves at the level of independent cells. The dynamic of such system is described by nonlinear differential equations.

3. The systems with nervous-system organization. They are based on the cardinal principles of human nervous system construction. As systems with reconversion of images, this is a dynamic system. The dynamic in such systems is described by complex nonlinear differential equations.

The main unit of all layers of the structure is a cell (informal neuron). The cell is considered to be a minimal element, capable to transform information. An informational model of the cell can be formally represented as the associative fuzzy-logical neural network with layered structure. Formality it can be as:

NM = {X, H_x, St, W, H_y, y}, ( 2 )

where X and y - input vector variables and single output variable; H_x and H_y - hidden variables for activator of the module; St and W - structure and connection weights for the activator of module. This model is executes any vector-vector mapping X ® y on many examples (patterns) with learning.

This informational model is supported by the functional model of the cell

C = (IG, A(a₁,…, a_n, S, W), DG, T_C), ( 3 )

where IG - informational granulator forming the set H_x; A(*) - activator consisting of the set of activator's elements a_i (i=1,…,N); these elements are doing transformation of the informational granules and composing the set H_y according to the links S and weights W; DG - degranulator, which form the output parameter y; T_C - adjuster composing the collection of the links S and weights W under circumstances of the cell tuning for the proper information map X ® y.

In according to this formalization the cell accepts and transforms information at granular mode. That's means that cell compares the current value of the input parameter to a priory defined granules of the parameter. The granule's boundaries can be implicit, fuzzy or random. The estimations of the parameters on suitability input values to granules are transformed in the activator. Activator is the layered associative structure composed of the activator elements, having such basis functions as unity, intersection, inversion and weigh. Using that complete set of operations, it's possible to store associatively and realize any input-output maps.

We can introduce three models of the activator [3].

1. Projection model of the activator. It’s the simplest from all developed structures.

, (4)

where n – the number of inputs; w_ij –weight coefficient; k_i – the number of terms on i-th input; l – index corresponding to the first term of the r-th output; p – index corresponding to the first term of the i-th input.

The most effective way of projection model learning is application of a genetic algorithm. The process of the learning consists in tuning of weight coefficients.

This model has serious limitations in the range of application. It can’t be learned to some kinds of image transformation where full information can’t be stored in projections.

Here are the advantages of the structure: (i) low requirements to available memory; (ii) simplicity of the structure; (iii) high ability to generalization.

Here are disadvantages of the structure: (i) applicability to a limited class of problems, in which projections allow to keep full information about input-output transformation; (ii) impossibility of using simple learning algorithms.

2. Grid model of the activator. This model is universal as regards to ability to realize almost any input-output image transformation.

(5)

where n – the number of inputs; w_i…j – weight coefficient; m_i – the number of terms on i-th input. For learning can be used simple algorithm of weight correction.

Here are advantages of the structure: (i) simplicity of the structure and learning algorithm; (ii) versatility of structure, in the sense of abilities to realize practically any continuous function.

Here are disadvantages of the structure: (i) redundancy of internal structure and connected with this high requirements, which object presents to the volume of the memory; (ii) low ability to generalization.

, (6)

where n – the number of inputs; m – the number of examples; w_i – weight coefficient corresponding to i-th term; k – number of the defining term.

Adjusting of weight coefficients and coordinates and defining bounds and centers of terms are realized during the learning of the activator.

Here are advantages of the structure: (i) simplicity of the structure and learning algorithm; (ii) versatility of structure, in the sense of abilities to realize practically any continuous function; (iii) high ability to generalization; (iiii) low requirements to available memory.

Here are disadvantages of the structure: (i) some difficulties for hardware realization.

The cluster model is the most suitable for using in the CCS. This model combines advantages of grid and projection systems and unites versatility with low memory requirements.

In order to work properly the topology of the CCS must be organized according to preliminarily defined goal of the control in a defined environment. By analogy with natural systems the main principle of topology forming is a self-organization.

The self-organization is a complex process including self-assemblage, self-learning and self-update of the system.

Self-assemblage is the first stage of a self-organization. It’s a search procedure in the space of possible elements and the links between them.

Self-learning is an optimization procedure. It consists in tuning structure parameters and functions in order to provide adequate behavior of CCS in a continuously changing environment.

Self-update is a process of updating genetic information accumulated in CCS. System uses this information at a self-assemblage and a self-learning phases to direct searching and optimizing procedures.

The process of self-organization can be realized it three ways.

1. Centralized organization. In such case system is organized by single module containing meta-information about the whole system structure and monitors current fitness of the system in terms of the goal. Central module defines the behavior of the whole system.
2. Decentralized organization. In this case each cell holds information about its neighborhood and has access to the current fitness of the whole structure. Each cell defines its own behavior.
3. Combined organization. It’s a combination of centralized and decentralized organizations. In this case each cell contains information about its neighborhood and defines its behavior. But in addition system has central unit which holds meta- information and in some cases can dictate a behavior to any cell.

The main purpose of CCS is a multi-channel complex synergetic control, similar to that, which is provided by central nervous system of a human. The range of application of CCS is extremely wide. It varies from synergetic control of an artificial hand, where each finger must work in according to others to provide an execution of a common task, to application in expert systems for forecasting of stock prices.

1. Stankevich L.A., Timakin D.L. Fuzzy-logical cognition control systems of the dynamical objects. //Proceeding of the Third International Symposium INTELS’98, Pskov 1998 (in Russian).
2. Stankevich L.A. A cognitive agent for soccer game // Proceeding of the First Workshop of Central and Eastern Europe on Multi-agent Systems CEEMAS'99 (S-Petersburg, Russia, June, 1999). Printed by "Anatolia", S-Petersburg, 1999.
3. Stankevich L.A., Timakin D.L. at al. Fuzzy-logical self-learning neural-network controller. // Proceeding of the Fifth Joint Russian conference "Neurocomputers and their application. ". Moscow, 17-19 February 1999.