Computational Thinking - Учебное пособие, страница 6

What is the problem of feedforward net implementationand how to overcome it?7. How is the training algorithm for feedforward netimplemented?8. What are the parts of a genetic algorithm?5. Translate from Russian into English:Для того, чтобы нейронная сеть могла решать задачи, еенеобходимо обучить. Для искусственных нейронныхсетей под обучением понимается процесс настройкиархитектуры сети (структуры связи между нейронами) ивесов семантическихсвязей (влияющих не сигналыкоэффициентов)дляэффективногорешенияпоставленной задачи.

Обычно обучение нейронной сетиосуществляется на некоторой выборке. По мере процессаобучения, который происходит по некоторому алгоритму,сеть должна все лучше (правильнее) реагировать навходные сигналы.30Оказывается, что после многократного предъявленияпримеров веса нейронной сети стабилизируются, причемнейронная сеть дает правильные ответы на все (илипочти все) примеры из базы данных. В таком случаеговорят, что нейронная сеть обучена. В программныхреализациях можно видеть, что в процессе обучениявеличина ошибки постепенно уменьшается.

Когдавеличина ошибки достигает нуля или приемлемогомалого уровня, тренировку останавливают, а полученнуюнейронную сеть считают натренированной и готовой кприменению на новых данных. Степень достаточностиобучения нейронной сети во многом определяетсяисходя из условий конкретной задачи и желаемогорезультата.6. Summarize the text and express your own opinion. Hereare some possible statements to support:1. In general, model predictive control is a strategy which isbased on the explicit use of some kind of process model topredict the controlled variables over a certain time horizon,the prediction horizon.2.

The prediction error between the measured process outputand the predicted output is used for disturbance estimation.3. For the prediction of the behavior of, for instance, theneutralization reactor, a feedforward network with sigmoidactivation function is chosen.4. The neural network consists of four layers.5. The network topology was chosen based on experimentswith different structures.6. The prediction error is not very sensitive to the number ofneurons in the hidden layer.31Biologically Inspired ComputingWords and phrases:RelentlessTo inspireTo conceiveTo emulateMomentumPermutationA cellular automatonSwarm intelligenceAmorphous computingTo straddleUnsupervised learningForagingTo exhibitSelf-replicationNested hierarchyTo delimitWhereinA strandCut-and-pasteTweezersEntanglementLogic gatesTo sustainA toolkitA motifTo yieldA ciliate– безжалостный, неослабевающий- вдохновлять- задумывать, замышлять- имимтировать, моделировать- толчок, движущая сила- преобразование, перестановка- матричный, клеточный автомат- роевый интеллект- хаотичныевычисления- охватывать- неконтролируемое обучение- добывание пищи- обнаруживать, проявлять- самовоспроизведение- иерархия с вложениями- разграничивать, ограничивать- при котором- цепочка, нить, составляющая- вырезание и вставка- пинцет- сцепленность, спутанность- логические элементы- подкреплять, поддерживать- набор инструментальных средств- повторяющийся узор- предоставлять- инфузорияDespite the relentless, breathtaking advances in computingand related technologies, we continue to be humbled by thevariety, adaptability, and sophistication of the natural worldaround us.

From the beginning, a divide has grown betweencomputational scientists on whether to continue creatingfaster, more efficient algorithms and hardware that exhibit32centralized control or to place less emphasis on speed andefficiency than on robustness, adaptability, and emergentorganization from the interaction of many loosely coupledprocesses. These latter approaches have come to be known asbiologically inspired computing, which is a field that linksvarious disciplines such as artificial intelligence, evolutionarycomputation, biorobotics, artificial life, and agent-basedsystems.Natural computing investigates models and computationaltechniques inspired by nature and, dually, attempts tounderstand the world around us in terms of informationprocessing.

It is a highly interdisciplinary field that connectsthe natural sciences with computing science, both at the levelof information technology and at the level of fundamentalresearch. As a matter of fact, natural computing areas andtopics include pure theoretical research, algorithms andsoftware applications, as well as biology, chemistry andphysics experimental laboratory research. A closer look atnature from the point of view of information processing canand will change what we mean by computation.Nature as InspirationAmong the oldest examples of nature-inspired models ofcomputation are the cellular automata conceived by Ulamand von Neumann in the 1940s.

Cellular automata have beenapplied to the study of phenomena as diverse ascommunication,computation,construction,growth,reproduction, competition, and evolution. Cellular automatahave been extensively studied as an alternative explanation tothe phenomenon of emergence of complexity in the naturalworld, and used, among others, for modeling in physics, andbiology.In parallel to early comparisons between computing machinesand the human nervous system, McCulloch and Pittsproposed the first model of artificial neurons.

This researcheventually gave rise to the field of neural computation, braintheory or computational neuroscience, and artificial neuralnetworks or simply neural networks.33While Turing and von Neumann dreamt of understanding thebrain, and possibly designing an intelligent computer thatworks like the brain, evolutionary computation emerged asanother computation paradigm that drew its inspiration fromDarwinian evolution.

Rather than emulating features of asingle biological organism, evolutionary computation draws itsinspiration from the dynamics of an entire species oforganisms. An artificial evolutionary system is acomputational system based on the notion of simulatedevolution.Evolutionary systems gained momentum in the 1960s, andevolved along three main branches: evolution strategies todayare used for real-valued as well as discrete and mixed types ofparameters; evolutionary programming is used today for realvalued parameter optimization problems; genetic algorithmsare today modified heavily for applications to real-valuedparameter optimization problems as well as many types ofcombinatorial tasks such as, for example, permutation-basedproblems.Cellular automata, neural computation, and evolutionarycomputation are the most established “classical” areas ofnatural computing.

Several other bio-inspired paradigmsemerged more recently, among them swarm intelligence,artificial immune systems, artificial life, membranecomputing, and amorphous computing.A computational paradigm straddling at times evolutionarycomputation and neural computation is swarm intelligence. Aswarm is a group of mobile biological organisms (such asbacteria, ants, termites, bees, spiders, fish, birds) whereineach individual communicates with others either directly orindirectly by acting on its local environment.

Theseinteractions contribute to distributive collective problemsolving. Swarm intelligence, sometimes referred to ascollective intelligence, is defined as the problem-solvingbehavior that emerges from the interaction of such acollection of individual agents.Particle swarm optimization was introduced as a new34approach to optimization that had developed from simplemodels of social interactions, as well as of flocking behaviorin birds and other organisms.

A particle swarm optimizationalgorithm starts with a swarm of “particles”, each representinga potential solution to a problem, similar to the population ofindividuals in evolutionary computation. Particle SwarmOptimization algorithms have been used to solve variousoptimization problems, and have been applied to unsupervisedlearning, game learning, scheduling and planning applications,and design applications.Ant algorithms were introduced to model the foragingbehavior of ant colonies. In finding the best path betweentheir nest and a source of food, ants rely on indirectcommunication by laying a pheromone trail on the way backto the nest if they found food, and following theconcentration of pheromones in the environment if they arelooking for food.

This foraging behavior has inspired a largenumber of ant algorithms used to solve mainly combinatorialoptimization problems defined over discrete search spaces.Bio-inspired algorithms can exhibit strength throughflexibility, or strength in numbers: they often work well evenwhen the desired task is poorely defined, adapt to unforeseenchanges in the task environment, or achieve global behaviorthrough interaction among many, simply programmed agents.Artificial immune systems are computational systems devisedstarting in the late 1980s and early 1990s as computationallyinteresting abstractions of the natural immune system ofbiological organisms. Viewed as an information processingsystem, the immune system performs many complexcomputations in a highly parallel and distributed fashion. Ituses learning, memory, associative retrieval, and othermechanisms to solve recognition and classification problemssuch as distinction between self and nonself cells, andneutralization of nonself pathogenic agents.

Indeed, thenatural immune system has sometimes been called the“second brain”, because of its powerful informationprocessing capabilities.35The applications are varied and include computer virusdetection, anomaly detection in a time series of data, faultdiagnosis,patternrecognition,machinelearning,bioinformatics, optimization, robotics, and control. Recentresearch in immunology develops what is known as the“danger theory”, wherein it is believed that the immunesystem differentiates between dangerous and non-dangerousentities. These ideas have started to be exploited in artificialimmune systems in the context of computer security.Artificial life (ALife) attempts to understand the very essenceof what it means to be alive by building ab initio, within insilico computers and other “artificial” media, artificialsystems that exhibit properties normally associated only withliving organisms.

Building on the ideas of evolutionarycomputation, the pioneers of artificial life experimented withevolving populations of “artificial creatures” in simulatedenvironments. These ideas were taken one step further, bycombining the computational and experimental approaches,and using rapid manufacturing technology to fabricatephysical robots that were materializations of their virtuallyevolved computational counterparts.