Pages:     | 1 |   ...   | 51 | 52 || 54 | 55 |   ...   | 58 |

state to extend the service life. Generation and Step 1. A teaching pattern is made on the base development of more effective, quite simple of known information. A neuro-network is and accessible methods of structural behaviour taught (selection of network architecture, forecast represents a practical interest for which is the best to take into account a design engineers, as an auxiliary procedure for problem specificity, is made). A forecasting design and investigation, for computationallydirection is chosen.

intensive and expensive actual tests and Step 2. A pattern thickening near the border of experiments.

known zone by means of network functioning As an example of effective offered neuro(interpolation) and network teaching on the net forecasting method application we apply base of enlarged pattern with the use of computational investigation data of timeprevious synaps chart with previous or dependent behavior of a cylindrical shell (Fig.

improved architecture.

2) with framework of glued wood and a plank Step 3. Making of short forecasting step, i.e.

layer under constant load [7]. Conventional an extrapolation over rather short interval (no construction working time was set to be greater then data step in the teaching pattern).

years and divided into 7 equal time intervals.

Step 4. Adding an obtained forecast to the At the end of each interval the stress and strain pattern, and neuro-network teaching on the base of an enlarged pattern. A further XVI state was defined, the strain integral module in can be realized by elaborated method (in essential elements was improved. contrast to other well-known forecasting The use of neuro-net forecasting showed methods). The use of neuro-net forecasting in that it is possible to be limited to only 3 this process can allow considerably reduce corrective actions of modulus of deformation, testing time and materials cost [5-8].

which are attended by series of complex Conclusion nonlinear analysis of the given shell. The other Evaluating neuro-net forecasting four corrective actions to make with the aid effectiveness, we should point out along with of neuro-net forecasting.

advantages objective delimitations. Primarily, these are a necessity of system-defined estimation of initial information ampleness and reliability, a choice of effective information organization and forecasting trend in multidimensional and multiparameter problems, etc. Such delimitations are inherent to each cognitive process and should take a proper place in forecasting results estimation.

Fig. 2. The under test cylindrical shell scheme References Two kinds of neuro-net forecasting were 1. V.G. Redko. Evolutional cybernetics. Lectures carried out traditional neuro-extrapolation on neuro-informatics, IV Russ. Conf and stepwise neuro-forecast with teaching and Neuroinformatica2002. Part 1 M.: MIFI, 2002, 104 pp. (in Russian) development of neuromodel for every forecast 2. G.G. Malinetskiy, S.P. Kurdyumov Nonlinear stage. Four groups of case study (initial stage dynamics and forecasting problems. Bulletin of and three next steps of deformation modulus Russian Academy of Sciences, 2001, Vol 71, 3, correcting) were presented for training P..210-232. (in Russian) network. Five most representative data were 3. Z. P. Szewczyk. Neural Network Based Extrapolation Strategies in Structural Analysis and selected to be the input data (modulus of Design // Struct.Optim.(Germ.) P.238-255 (1999).

deformation E for all subsystems). The output 4. Abovskiy N.P., Deruga A.P., Maximova O.M., data (16 in all) are data of nonlinear SSS Svetashkov P.A. Neuro-Control Structures and Systems analysis for all regions of test construction.

/ M.: Radiotechnica, 2003. 368p. (Book 13 of Scientific The choice of the most effective neural net Series Neuro-computers and their application, edit.

.I. Galushkin). (in Russian) for forecasting was made. RBF (with radial5. Abovskiy, N. P. Neuro-Prognosis Based on basis neuron function) became it. Network Step Model with Teaching for Natural Tests Results of teaching at every step was carried out by Back Building Structures / N. P. Abovskiy, . .

Propagation, Quick Propagation and QuasiMaximova // j. Optical Memory & Neural Networks Newton methods.

(Information Optics), 2007, Vol. 16, No.1, pp. 40-46. // j. SPRINGER, 2007.

In accordance with computation results 6. Maximova O.M. Creation and Application of precision of the step-by-step neuro-net Neuro-net Technologies for forecasting in Building forecasting is considerably better than of Structures and Building Mechanics / Fundamental and traditional (one step) neuro-extrapolation.

Applied Science Problems. Vol. 2. - 1-st Intern.

Even at the latest stage (50 years) a step-by- Symposium Works. - .:RAS, 2010.P.3-24. (in Russian) step forecast error did not exceed 3.8% versus 7. Maximova O.M. Effective step-by-step 11.07% traditional extrapolation error.

method of neuro-net forecasting for research tasks and In the course of carrying out in-place tests structural design and their elements // address to IV we do not always manage to lead construction Iintern. Pract. Conf. Theory and practice of a building, to destruction. As in-place test of engineering construction and elements of structure proportioning.

Analytic and numerical methods. 29 June, 2011.

structures is a long-continued, labourMoscow, MICE-MSBU. 9p. (in Russian) consuming and expensive process, only few 8. Maximova O.M. Neuro-net forecasting test steps are often realized. On the other hand technologies for dynamic problems of engineering there is a need to know (to forecast) behaviour structures. // Scientific Session MIFI - 2012. IV Russ.

of the structure on the next stage, i.e. to carry Conf Neuroinformatica-2012 // Coll. Scientific Works in 3 parts. Part 2. M.: MIFI, 2012. P.71-82. (in Russian) out quite deep forecasting. Such possibility 4- ۻ BIOLOGICALLY INSPIRED COGNITIVE ARCHITECTURES (BICA) A. V. Samsonovich Krasnow Institute for Advanced Study, George Mason University, Fairfax, Virginia, USA alexei@bicasociety.org Biologically inspired cognitive architectures (BICA) are considered biologically inspired when it is a powerful emergent approach in cognitive modeling based on the principles of natural intelligent and intelligent agent design, bridging the boundaries systems, primarily, the human brain-mind.

between artificial intelligence, cognitive science, and The idea of the BICA approach proves neuroscience. This work reviews its key ideas and powerful, because cognitive, metacognitive challenges. It is discussed how through a variety of problems and solutions, from navigation to emotion and and learning capabilities of natural biological goal generation, BICA will allow us to make progress systems are characterized by a high degree of toward building a computational equivalent of the robustness, flexibility, sustainability and human mind.

adaptability to real-life situations.

The BICA approach overcomes many Introduction limitations of traditional approaches. E.g., unlike neural modeling, it does not impose The acronym BICA stands for hard biological constraints. Another limitation biologically inspired cognitive architectures. It of realistic neuronal networks is their was coined by the Defense Advanced conceptual ceiling that was reached in Research Projects Agency (DARPA) in 2005, connectionism several decades ago: these as the name of one of its most advanced models still cannot combine the desired degree programs intended to develop psychologically of integration and completeness with the and neurobiologically based computational human level of cognition. Another example of models of human cognition. The program was a problem in non-BICA artificial intelligence terminated by the US Congress in 2006. Since approaches is the tremendous amount of then, many new BICA-related projects and human labor required for implementing a initiatives were funded by DARPA and other capability that may be laser-focused on a funding agencies around the world, under specific task and, as a result, brittle [15, 16].

various names that are not always obviously The BICA approach makes a significant related to each other.

step forward with respect to its predecessors.

The emergent field of BICA research It starts at a higher level of abstraction brings together the old great goals of artificial compared to neural modeling: at a cognitiveintelligence [1], cognitive [2] and neural system level, with elements of the model modeling under a new umbrella: an emergent having semantics attributed to them. This overarching BICA Challenge [3, 4]. In short, allows one to implement higher-level primary the challenge is to create a computational concepts in the architecture at the time of its equivalent of the human mind.

design rather than wait for them to emerge Since the onset of cognitive modeling as through self-organization. At the same time, a research paradigm, attempts are made to most cognitive skills and knowledge are implement and study complete cognitive expected to develop naturally in the agent agents embedded in virtual or physical through various forms of learning, instead of environments [2]. Models of this sort are manual programming. As a consequence, a known as cognitive architectures [5-9]. More BICA agent can start its life as a minimal, precisely, the term cognitive architecture is domain-independent cognitive embryo. In understood as a computational framework for addition, the biological fidelity of BICA designing a complete intelligent agent [2, 9], makes them virtually human-compatible.

and not only its architecture. The most popular In the nutshell, these are the main cognitive architectures are Soar [10-12] and features that make BICA stand above other, ACT-R [13, 14]. A cognitive architecture is XVI more traditional approaches in artificial models like semantic cognitive maps [18], intelligence, cognitive and neural modeling. semantic networks, belief networks and graphical models, etc.

The Four Pillars of BICA One illustrative example of a BICA component is the grid-cell-network model of The foundations of BICA include: (1) representations of space. Grid cells were computational neuroscience, (2) cognitive recently discovered in the entorhinal cortex in modeling, (3) theoretical functional models of rodents, and also in humans [21]. Typically consciousness and the Self, and (4) artificial combined with a model network of the intelligence. All these fields have different hippocampal place cells, this component is goals: (1) and (2) aim to understand how the useful for spatial learning and navigation.

brain works at the level of neurons and at the level of behavior and underlying cognitive Basic Theory and Methods of BICA processes, (3) aims to understand the nature and mechanisms of the human mind and to A generic BICA (Figure 2) includes mimic them in artifacts, and finally, (4) functional components that support basic pursues practical goals regardless of the nature forms of human cognition. By tradition, most of solutions. BICA is a unification of the four of these components are regarded as memory research paradigms that reconciles their goals systems. Interpretation of the data on human and amplifies their individual strengths. This memory systems borrowed from cognitive central idea of integration works through a neuropsychology [17] by computer scientists variety of problems and solutions, from remains highly controversial: the terminology navigation to emotion and goal generation, is frequently inconsistent across domains.

showing how BICA will allow us to make progress toward solving the BICA Challenge.

Figure 1. Building blocks of BICA representing the four pillars: A, a schema, B, a semantic space, C, a neural Figure 2. Main components of a generic BICA.

network, D, a hard-coded sequence of actions.

But on the other hand, terms used in computer Building blocks of BICA include science are usually defined very precisely and elements of various nature (Figure 1): e.g., constructively, unlike in psychology.

schemas representing symbolic knowledge, Elements of memory representations have neural nets and semantic spaces representing different names in different cognitive neuromorphic and analog components, and architectures: schemas, chunks, states, etc.

hard-coded algorithmic solutions of specific The term schema is used here.

tasks, including signal-to-symbol conversion Main memory systems in a generic BICA and motion control. As a result, many BICA can be briefly characterized as follows below.

are hybrid architectures: they combine They include procedural, working, semantic neuromorphic (neural-network), symbolic, and episodic memory systems.

algorithmic and analog elements, including 4- ۻ Semantic memory consists of schemas that may be organized into a semantic net. In addition to this organization, schemas may be allocated as points in an


semantic space based on their semantics, forming a semantic cognitive map [18].

Procedural memory consists of primitives that subserve specialized cognitive functions and skills: e.g., interface with the environment.

Working memory consists of active Figure 3. The hierarchy of cognitive capabilities that schemas that are bound to each other and in constitute parts of the BICA Challenge (based on [4]).

addition may be clustered into mental states [22].

Implementation of these critical Episodic memory consists of groups of capabilities at a human level of abstraction frozen mental states that were previously would be necessary for overcoming barriers active (i.e., were present in working memory) on the way to solving the BICA Challenge.

The notion of episodic memory includes not Some of them are already available in only retrospective memories of actual implemented BICA, others are not experiences attributed to the self [19], but also implemented or may only seem to be prospective memories, including plans [20] implemented.

and imaginary experiences (dreams).

Pages:     | 1 |   ...   | 51 | 52 || 54 | 55 |   ...   | 58 |

2011 www.dissers.ru -

, .
, , , , 1-2 .