Genotype dynamic for agent neuroevolution in artificial life model

Cooperation behavior is one of the most used and spread Multi-agent system feature. In some cases emergence of this behaviour can be characterized by division of population on co-evolving subpopulations [1], [2]. Group interaction can take not only antagonistic conflict form but also genetic drift t...

Повний опис

Збережено в:

Бібліографічні деталі
Дата:	2017
Автори:	Zavertanyy, V., Makarenko, A.
Формат:	Стаття
Мова:	English
Опубліковано:	Навчально-науковий комплекс "Інститут прикладного системного аналізу" НТУУ "КПІ" МОН та НАН України 2017
Назва видання:	Системні дослідження та інформаційні технології
Теми:	Математичні методи, моделі, проблеми і технології дослідження складних систем
Онлайн доступ:	http://dspace.nbuv.gov.ua/handle/123456789/151065
Теги:	Додати тег Немає тегів, Будьте першим, хто поставить тег для цього запису!
Назва журналу:	Digital Library of Periodicals of National Academy of Sciences of Ukraine
Цитувати:	Genotype dynamic for agent neuroevolution in artificial life model / V. Zavertanyy, A. Makarenko // Системні дослідження та інформаційні технології. — 2017. — № 1. — С. 75-87. — Бібліогр.: 22 назв. — англ.

Репозитарії

Digital Library of Periodicals of National Academy of Sciences of Ukraine

id	irk-123456789-151065
record_format	dspace
spelling	irk-123456789-1510652019-04-24T01:25:41Z Genotype dynamic for agent neuroevolution in artificial life model Zavertanyy, V. Makarenko, A. Математичні методи, моделі, проблеми і технології дослідження складних систем Cooperation behavior is one of the most used and spread Multi-agent system feature. In some cases emergence of this behaviour can be characterized by division of population on co-evolving subpopulations [1], [2]. Group interaction can take not only antagonistic conflict form but also genetic drift that results with strategies competition and assimilation [3]. In this work we demonstrate different relation between agent grouping and they behavior strategies. We use approach proposed in work [2] methodology of agent genotype dynamic tracking, due to this approach the evolving population can be presented in genotype space as a cloud of points where each point corresponds to one individual. In current work consider the movement of population centroid – the center of the genotype cloud. Analysis of such trajectories can shad the light on the regimes of population existence and genesis. Кооперативна поведінка є однією з найбільш часто використовуваних та поширених рис для багатоагентних систем. У деяких випадках поява такої поведінки пов’язана із поділом населення на співіснуючі субпопуляції [1, 2]. Групова взаємодія може набувати не лише форми антагоністичного конфлікту, але й зумовлюватися генетичним дрейфом, який приводить до конкуренції поведінкових стратегій та можливої асиміляції [3]. Продемонстровано різні види залежностей між групами агентів та їх поведінковими стратегіями. Використано методологію спостереження за динамікою агентного генотипу [2], відповідно до якої популяція у просторі генотипів може мати вигляд хмари точок, кожна точка якої відповідає одній особині. Розглянуто динаміку центроїда населення — центра хмари генотипу. Аналіз таких траєкторій може сприяти дослідженню різних режимів існування популяції та їх зародження. Кооперативное поведение является одной из наиболее часто используемых и распространенных черт для многоагентных систем. В некоторых случаях появление такого поведения связано с разделением населения на сосуществующие субпопуляции [1, 2]. Групповое взаимодействие может принимать не только форму антагонистического конфликта, но и обуслoвливаться генетическим дрейфом, приводящим к конкуренции поведенческих стратегий и возможной ассимиляции [3]. Продемонстрированы различные виды зависимостей между группами агентов и их поведенческими стратегиями. Использована методология наблюдения за динамикой агентного генотипа [2], согласно которой популяция может быть представлена в пространстве генотипов в виде облака точек, где каждая точка соответствует одной особи. Рассмотрена динамика центроида популяции — центр облака генотипа. Анализ таких траекторий может помочь исследованию различных режимов существования популяции и их зарождения. 2017 Article Genotype dynamic for agent neuroevolution in artificial life model / V. Zavertanyy, A. Makarenko // Системні дослідження та інформаційні технології. — 2017. — № 1. — С. 75-87. — Бібліогр.: 22 назв. — англ. 1681–6048 DOI: https://doi.org/10.20535/SRIT.2308-8893.2017.4.06 http://dspace.nbuv.gov.ua/handle/123456789/151065 518.58:519.2:504 en Системні дослідження та інформаційні технології Навчально-науковий комплекс "Інститут прикладного системного аналізу" НТУУ "КПІ" МОН та НАН України
institution	Digital Library of Periodicals of National Academy of Sciences of Ukraine
collection	DSpace DC
language	English
topic	Математичні методи, моделі, проблеми і технології дослідження складних систем Математичні методи, моделі, проблеми і технології дослідження складних систем
spellingShingle	Математичні методи, моделі, проблеми і технології дослідження складних систем Математичні методи, моделі, проблеми і технології дослідження складних систем Zavertanyy, V. Makarenko, A. Genotype dynamic for agent neuroevolution in artificial life model Системні дослідження та інформаційні технології
description	Cooperation behavior is one of the most used and spread Multi-agent system feature. In some cases emergence of this behaviour can be characterized by division of population on co-evolving subpopulations [1], [2]. Group interaction can take not only antagonistic conflict form but also genetic drift that results with strategies competition and assimilation [3]. In this work we demonstrate different relation between agent grouping and they behavior strategies. We use approach proposed in work [2] methodology of agent genotype dynamic tracking, due to this approach the evolving population can be presented in genotype space as a cloud of points where each point corresponds to one individual. In current work consider the movement of population centroid – the center of the genotype cloud. Analysis of such trajectories can shad the light on the regimes of population existence and genesis.
format	Article
author	Zavertanyy, V. Makarenko, A.
author_facet	Zavertanyy, V. Makarenko, A.
author_sort	Zavertanyy, V.
title	Genotype dynamic for agent neuroevolution in artificial life model
title_short	Genotype dynamic for agent neuroevolution in artificial life model
title_full	Genotype dynamic for agent neuroevolution in artificial life model
title_fullStr	Genotype dynamic for agent neuroevolution in artificial life model
title_full_unstemmed	Genotype dynamic for agent neuroevolution in artificial life model
title_sort	genotype dynamic for agent neuroevolution in artificial life model
publisher	Навчально-науковий комплекс "Інститут прикладного системного аналізу" НТУУ "КПІ" МОН та НАН України
publishDate	2017
topic_facet	Математичні методи, моделі, проблеми і технології дослідження складних систем
url	http://dspace.nbuv.gov.ua/handle/123456789/151065
citation_txt	Genotype dynamic for agent neuroevolution in artificial life model / V. Zavertanyy, A. Makarenko // Системні дослідження та інформаційні технології. — 2017. — № 1. — С. 75-87. — Бібліогр.: 22 назв. — англ.
series	Системні дослідження та інформаційні технології
work_keys_str_mv	AT zavertanyyv genotypedynamicforagentneuroevolutioninartificiallifemodel AT makarenkoa genotypedynamicforagentneuroevolutioninartificiallifemodel
first_indexed	2025-07-13T00:58:56Z
last_indexed	2025-07-13T00:58:56Z
_version_	1837491385060556800
fulltext	© V. Zavertanyy, A. Makarenko, 2017 Системні дослідження та інформаційні технології, 2017, № 1 75 TIДC МАТЕМАТИЧНІ МЕТОДИ, МОДЕЛІ, ПРОБЛЕМИ І ТЕХНОЛОГІЇ ДОСЛІДЖЕННЯ СКЛАДНИХ СИСТЕМ УДК 518.58:519.2:504 DOI: 10.20535/SRIT.2308-8893.2017.4.06 GENOTYPE DYNAMIC FOR AGENT NEUROEVOLUTION IN ARTIFICIAL LIFE MODEL VALENTINE ZAVERTANYY, ALEKSANDR MAKARENKO Abstract. Cooperation behavior is one of the most used and spread Multi-agent system feature. In some cases emergence of this behaviour can be characterized by division of population on co-evolving subpopulations [1], [2]. Group interaction can take not only antagonistic conflict form but also genetic drift that results with strategies competition and assimilation [3]. In this work we demonstrate different relation between agent grouping and they behavior strategies. We use approach proposed in work [2] methodology of agent genotype dynamic tracking, due to this approach the evolving population can be presented in genotype space as a cloud of points where each point corresponds to one individual. In current work consider the movement of population centroid – the center of the genotype cloud. Analysis of such trajectories can shad the light on the regimes of population existence and genesis. Key words: artificial life, multiagent systems, neuroevolution INTRODUCTION Artificial Life (Alife) is an interdisciplinary research field, which try to investi- gate and use the properties of living systems or systems which include a large number of living components (for example, individuals). Alife usually brings to- gether biologists, philosophers, physicists, computer scientists, chemists, mathe- maticians, artists, engineers, and more. The examples of Alife fields are numerous and includes artificial (digital) ecosystems, artificial society, evolutionary robot- ics, biology, origin of life — see for examples in [4, 11], and many others. Alife systems are implemented as software and as hardware (see for recent review [10], [12]). Remark that one of the important examples of the software Alife studies build and explore digital ecosystems that provide novel methods to study evolu- tion. These studies can be useful in answering questions about laws how evolution works and how to operate it. Traditional evolution in real biological systems is extremely slow to study. The computation Alife aims to put the evolution process into action on a computer so time for evolution to go on is only limited by proces- sor performance. Embracing evolution instruments opens opportunities for re- searching a great variety of problems that are linked with it. Artificial evolving systems are used to build complex systems that expose intellectual behavior and study the link between intellectuality and complexity [13]. Alife systems are plau- V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 76 sible playground to explore the mechanisms of adaptation: general evolving sys- tem features such as speciation ([2], [13]), aging ([14]), cooperation ([15]), devel- opmental processes in artificial systems [16], and learning. Many models are developed in purpose to study social, ecological, swarm- ing, artificial life and other topics. Despite the progress of other models, the inter- connection between genotype and phenotype dynamic is still quite an unexplored issue; in current study we reveal an example of such unclearness that lurks in dy- namic of the system. By one of the goals of the study, we want to concentrate on the more detailed research of agent phenotype sustainability and what it depends of. Further in this work, we discuss the dependency of combat interaction from input resource value and examine the sustainability of phenotypic assembly for- mation in homogeneous and heterogeneous spaces. These questions fit into the research field of Artificial Life determined by Bedau [11], and belong to a group of research areas that claim to: • determine predictability of evolutionary consequences of manipulating organisms and ecosystems; • determine minimal conditions for evolutionary transitions from specific to generic response systems; • determine what is inevitable in the open-ended evolution of life. Alife consolidate different research fields, such as, for example, hardware and software Alife. It could be used to study the evolution of complexity, robot- ics, and digital organisms. One of the main approach of constructing simulation models in Alife is multi-agent methodology that is broadly used in the study of complex adaptive systems. Individual-based approach surmounts difficulties of equation-based models by granting additional flexibility for both development and analysis of the model [12]. The popularity of multi-agent approach springs from early researches such as Sugarspace [2], Bugs [18], Echo [19] and Poly- world [17] models. One of pioneer models of Artificial Life is the model of bugs on spatial lattice that was proposed by Norman H. Packard [18] denotes the im- portance of shift from extrinsic to intrinsic adaptation approaches in the modeling of evolutionary processes. Packard proposed to change the point of view on fit- ness in models of biological systems. He claimed that extrinsic approach of adap- tation such that is defined by an a priori fitness function that assumes averaging of the environment and individual interactions could inflict limitations on the bio- sphere. Such limitation takes place, for the organism affects its environment and other organisms, altering the whole biosphere and eventually its own possibility to exist, i. e. its own fitness [18]. The author defines the intrinsic adaptation of a system as a process of changes in interactions of all parts of the system aiming to fit it and permanently changing the environment. As a result of first simulations of his model, H. Packard introduced the notion of an a posteriori fitness function for the intrinsic adaptation evolutionary process and demonstrated with its help the emergence of specific behavior that is inherent for some individuals. This change in the concept of adaptation shifts the focus to the emerging characteris- tics of the system that can be treated as an a posteriori fitness function. The ex- amples of such values could be population size over time, sustainability of emerg- ing phenotypic assemblies under different factors such as environmental changes or arm races and other system features. In particular work the size of agents’ group with common phenotype (behavior strategy) is treated as the a posteriori fitness function. Echo model is a Complex Adaptive System that was built with a purpose of extending genetic algorithms approach to ecological setting by adding geography Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 77 (location), competition for resources and interaction among individuals (coevolu- tion). The model itself is intended to study patterns of behavior that are how re- sources flow through different kinds of ecologies, how cooperation among agents can arise through evolution and arms races. Echo corresponds to a set of Echo models, in the system agents evolve empowered with combat, trade, move and mate abilities that are conditioned by their genotype and phenotype traits. Echo model consists of agents that are located in two-dimensional grid of sites, and each agent is located at a site, migration is supported. Many agents can occupy one site and there is a notion of neighborhood. The different kinds of resource randomly distributed between all cells. Agents use resource to pay metabolic tax and to perform trace, combat and mating actions. Reproduction can be sexual (crossover) and asexual (replication with mutation). The system study allows identifying parameters or collections of parameters that are critical for emergence of specific behavior, i.e., to perform sensitivity analysis [19]. Simulation results and their analysis allow scientists to build deep intuitions about how different as- pects of the digital ecosystem interact one another, reveal important dependen- cies, and provide understanding of how evolution interacts with ongoing dynam- ics of the ecosystem [19]. In continuation to work with Echo model family Hraber and Milne discov- ered the notion of the emergence of community assemblies, they showed the exis- tence of agent groups that share common behaviour that springs in order to re- sponse on interaction rules in model architecture [20]. Certain genotype assemblies (complementary genotypes) were born and formed quasi-stable domi- nation that was based on pairwise interaction between agents. In particular work we consider digital ecosystem with such emergent feature and show that changing of system property such as space heterogeneity contributes to sustainability of complementary phenotypic assemblies over time. By saying phenotypic assembly we consider group of agents that share similar behaviour. It should be noted that such assemblies are less complex than community assemblies presented in Hraber and Milne study because agents action portfolio in that model is wider: its agents can trade and mate in addition. While in particular model phenotypic assembly by definition not necessarily support internal group interactions. The further continuation of digital ecosystem models is the models where complex agent’s behaviour arises from the first principles: where it never was predefined by fitness function and emerges through adaptation process. Remark- able examples of such models are Michael Burtsev’s [2], [21] model and Robert Grass’ [13] model. One of the main achievements of their research is that agent speciation i. e. phenotypic grouping and distinction emerges without predefined fitness function. Agents occupy niches that expose predator, prey or even more sophisticated behaviour without extrinsic predisposition but as the result of the evolutional adaptation process. Michael Burtsev proposed a model that resembles pioneer Artificial Life’s Echo [19] and Bugs [18] models: the agents with simple behaviour are acting in a simple space. In the study [2] author develops latter models introducing kinship (by introducing culture affinity) and using the artificial neural network as a basis for agent’s actions. In this model no agent was given a predefined strategy, in- stead it emerge as phenotype feature from agent’s actions, defined by the neural network. By doing this, the author achieved a great variety of strategies that can take into account kinship of the object they interact with and are constructed from elementary actions as a result of evolution processes. Some of the strategies expose cooperative behaviour, where agents adjusted their behaviour due to V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 78 genotypic distance between each other. It was shown that in such model emerge strategies that correspond to those in well-known game theory - dove-hawk- bourgeois, where dove acts like peaceful harvester, hawk demonstrates aggressive behaviour attacking agents in neighborhood, and bourgeois that plays as dove when low on resource and displays hawk strategy in possession of it. Also, two new strategies of cooperative attack (when agent attack only non-relative ones) and defense (when agents gather in one location to defense themselves from aggression) were emerged [20]. The similar results with different model achieves research with novel artificial life model with predator-prey behavior in study [13], where agents are driven by fuzzy cognitive map. Considering results of artificial life modeling it can be concluded that such approach is not being controversial to game theory but on the contrary is an extension that provides new research hori- zons, such as finding evolutionary stable strategy, designing an open-ended evolution, exploring new sophisticated agent behavior, and analyzing system regularities, e. g. persistent emergence of group behavior and arm races. By studying the model, Burtsev proposed a novel methodology to categorize agents’ behaviour into strategies and to trace population genotype dynamic [2]. Author proposes alternative view on evolving systems that is inspired by dynamical systems theory. He points that the main notion in this theory is the trajectory and provides mechanisms of tracing the Artificial Life model development as if it moves along its own trajectory. An evolving population can be presented in genotype space as a cloud of points where each point corresponds to one individual. It is proposed to consider the movement of population centroid — the center of the cloud. Analysis of such trajectories shad the light on the regimes of population existence and genesis. Analysis of mentioned above researches of Alife models show that they open novel regularities and emergent behavior. Proceedings study of the similar models discovers new aspects of agents’ behavior dynamic can be studied. Evolution processes in the models of digital ecosystems are far from being clear and traceable, the interconnection between emergent features and system parameters are not yet properly established. In this work we study the dynamic of agent’s population genotype and phenotype using the novel methodologies from Burtsev’s work [2]. The aim of the work is to illustrate phenotype and genotype dynamic in continuous space artificial ecosystem model. We open the questions of genotype and phenotype transition and the importance of understanding and tracking the nature of such transitions. MODEL DESCRIPTION Agent. Common predator-prey model with continuous space was developed in the work. Agent percepts certain environmental variables such as: • resource allocated nearby; • current energy value that belongs to agent itself (r); • difference between maximum reachable (rMax = 500) and current energy values (r); • the most weaker and the most stronger nearby agent (affected by affinity mark: if agent is relative than input is multiplied with (-1)); • In reply to the input signals agent performs the following actions; • ‘rest’: stay on the current position paying the smallest energy price; • ‘eat’: start consumption of vegetation. Agent can consume defined vegetation value by one time; Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 79 • ‘move’: move towards current heading, if agent sees the vegetation he can change heading in order to reach it. Agent pays specified energy cost for movement; • ‘attack’: jump to victim agent nearby (hunter pays additional penalty for jump action) and attack him. Attacker pays fee to initiate aggression, if he is suc- cessful he gains all victim’s energy (energy consumption is limited by maximum reachable energy threshold) and victim dies; • ‘divide’: agent creates sibling nearby and gives to him a half of own en- ergy. Sibling receives parent’s genotype with mutations; • ‘escape’: if one agent wants to escape from another agent, he starts movement in opposite way from target agent with two times higher speed. Actions and theirs fee are listed in Table I. T a b l e 1 . Agent’s fee for actions Action Payment type Value Rest Once per action 15 Move Move forward By overcame path (velocitystep) 25 Divide Once per action 20 + r/2 Attack Escape Twice by overcame path (2velocity*step) 50 Agent’s behavior is determined by artificial neuronal network with one layer. Each agent’s sibling inherits neuronal matrix perturbed with some muta- tions after birth (action ‘divide’). Each agent is characterized by affinity marker: 3-dimentional vector which coordinates can take possible integer values in [–2, 2] interval. Agents are considered as relatives if Euclidean distance between theirs markers are less than 0.2 threshold. Probability to be succeed in attack is equals to ratio of victim’s and at- tacker’s accumulated energy. Agent’s actions are categorized and vector of agent strategies is generated using the methodology firstly presented in [2]: to show agent phenotype behavior, each agent was placed in hypothetical situation as if he interacts with other agent under various conditions i. e. agent’s internal energy indicator, agent’s relative affinity. Thus, agent is being stressed with six input test vectors and then strategy vector was generated according to his reaction (Table 2). For example, strategy ‘020202’ is typical hawk strategy: regardless of internal agent energy level, he will attack any stranger in his area and make no harm to relatives. See agent’s population dynamic in Fig. 1. Population downfall is caused by conflicts between different culture groups. T a b l e 2 . Vector of agent’s strategies. Where a є {0: “rest”, “eat vegetation”; 1: “escape”; 2: “attack”; 3: “divide”}, 4,3,2,1=i Low resource, max02,0 rr = Half of resource, max5,0 rr = Many resources, max98,0 rr = Relative behind Non-relative behind Relative behind Non-relative behind Relative behind Non-relative behind iA V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 80 Fi g. 1 . P op ul at io n ra te (“ Po pu la tio n_ co un t” c ur ve ) a nd a ge nt ’s st ra te gi es Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 81 Space. Space is continuous and 2-dimentional. Agent’s position is defined with real coordinates ),( yx . Space is logically divided on cells (Fig. 2), agent can overcome one cell by two time series with average speed. Agent can perform actions with object on distance closer than Action Radius. Agent is aware of all objects that are located in cells that bear to his cell. New vegetation is randomly appears in space with each time series. For heterogeneous spa- ce vegetation checks appears in some areas more fre- quently that in others (see Fig. 3). ANALYSIS Population Genotype Centroid. In addition to presenting phenotype as behavior strategies in current work method of tracking Population Genotype Centroid was used [2]. Consider genotype space G , its dimension equals to number of genes for each agent from population P (quantity of elements in agent’s matrix of neu- ral network — mnW × ). Let genotype of agent A be Gg∈ , each element of g : kg , mnk ,,1 K= corresponds to element of the matrix W : ijw , ni ,,1 K= , mj ,,1 K= . Thus, g defines the point in space G that corresponds to agent A . We should track its centroid pC to analyse the movement of the whole agent’s population cloud: ∑ N i ii g N =C 1 . Centroid trajectory can be maintained for a long time in some compact area (gravity field) of the space G , form closed curves corresponding to a specific period cycles, accidentally wander around space or jump from one area of attraction to Cell Length Action Radius Fig. 2. Agent’s space. Agent are presented with ar- rows, vegetation as checks Fig. 3. Example of the map of heterogeneous space V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 82 another. To identify patterns in these behaviors the following characteristic is used [2]: 2))()((),( tCTtC=Ttd i N i i −+∑ (1) Expression (1) gives the distance between the centroid locations at time t and Tt + [2]. Visualizing of dynamics of values ),( Ttd for different values of period T can get an idea of the character of centroid movement in space. For ex- ample, in case of random walk of centroid in space ),( Ttd will increase with in- creasing T (Fig. 4,a) If for some period of time centroid moves along a closed trajectory of period cT , then during this period of time the value of ),( cTtd will be close to zero. Other examples of such visualization shown in Fig. 4. In the context of the model considered in this paper genotype is a low-level rule for behavior — he abstract specification for agent — that then participates in local interaction of a large set of other types of behavior. Phenotype is the behav- ior patterns in time and space that develop from these non-linear, local interac- tions [2] (Fig. 5). a b c d Fig. 4. Example of the different centroid behavior dynamics and visualization of d(t,T) characteristic. Figures are taken from [2] 1 1 1 1 1 GTYPE PTYPE Evolution Local rules define simple nonlinear interaction on this level Global behaviour and patterns emerges on this level Fig. 5. Interaction of genotype (GTYPE) and (PTYPE) [8] Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 83 MAS BEHAVIOUR Mutli-agent system display variety of different scenarios of agent interactions: predator-prey cycles [2], [13], [5], competition between different behaviour strategies [1], [5], quasi-stable domination of certain strategy for a long period of time or intense variability of strategies. Let us show competitive interaction between behaviour strategies with simi- lar behavior. Refer to Fig 1 we can see that in some time series population dra- matically decreases (for 410, 700, and 900 thousands time series). Considering successful attack actions time series (Fig. 6) we can infer that aggressive competi- tion was taking place for this cases. Consider fist crisis episode: population decrease near 400 thousands time series (Fig. 7) — we can see coexistence of the almost similar strategies: “333020” and “333030”, the main and crucial difference between them is that first strategy provides aggressive actions to relative agents. Time passed leads to extinction of aggressive non-cooperative strategy. On the visualization of centroid dynamic (Fig. 8), we can see that all popula- Fig. 6. Successful attack actions count for experiment from Fig. 1 — Successful_Attack Fig. 7. Detailed fragment from Fig. 1 view. Competitive interaction between strategies with similar behaviour. “333020” — triangles, “333030” — circles 400.000 Fig. 8. Visualization of ),( Ttd cha- racteristic for centroid behavior dy- namics for experiment from Fig. 1 V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 84 tion crisis periods are accompanied with centroid transition from one area of at- traction to another (similar to Fig. 4,d dynamic). However, it is not clear enough whether centroid relocates from one state to another under aggression actions or under peaceful assimilation of agents, such cases were described other models [2]. Culture marker and genotype vector. Let us emphasize on interconnection between genotype vector and vector of relative affinity. Consider two experiments with various behaviour strategies dominate in different periods of time (Fig. 9 and Fig. 10). As we can see for the first experiment culture marker centroid dynamic is closely connected (Fig. 9,b and Fig. 9,c). Culture centroid even has additional transitions from one area of attraction to another. But, for the second experiment no significant culture marker centroid transitions were tracked unlike to genotype centroid dynamics (Fig. 10,b and Fig. 10,c). Such behaviour causes new questions like dependency of centroids transitions from culture transition between agents and culture marker dimension (as dimension was stated important parameter in [8]), formation of agent groups settlement in different areas and possible swarm- Fig. 9. Experiment with correlated transition of genotype and culture marker centroids: а — strategy series; b — genotype centroid transition dynamic visualization; c — marker of culture features dynamic visualization a b c Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 85 like behaviour, indicate and observe the crucial for transition components of agent’s genotype vector, provide clearness in stating of aggressive or peaceful (assimilative) centroid transition. DISCUSSION In the work we presented own variation of common model of digital ecology and demonstrated dynamics both of phenotype and genotype agent groups. The formation and further extinction of culture groups caused ether by aggressive predator-prey interaction or by competition of strategies with antagonistic or similar behavior. The following issues are important for further development of the model: • building new tools for analysis of agents population to enlighten how different interactions affect on group formation and its persistence; • determine influence of aggressive behavior and genetic transition to the change in the dominance of one strategy over the other; • analysis of agent group formation and its coevolution; • introduction of new types of interaction between agents. Fig. 10. Experiment where transition of genotype and culture marker centroids are not corresponded: a — strategy series; b — genotype centroid transition dynamic visualiza- tion; c — marker of culture features dynamic visualization a b c V. Zavertanyy, A. Makarenko ISSN 1681–6048 System Research & Information Technologies, 2017, № 1 86 REFERENCES 1. Kenneth Stanley O. Competitive coevolution through evolutionary complexification / Kenneth O. Stanley, Risto Miikkulainen // Journal of Artificial Intelligence Re- search. — 2004. 2. Burtsev М.S. Research into new types of self-organization and behavioral strategies: phd dis. / М.S. Burtsev // IAM M.V. Keldysh RAS. M. — 2005. — 120 p. 3. Epstein Joshua M. Growing artificial societies: social science from the bottom up / Joshua M. Epstein, R. Axtell. — Brookings Institution Press, 1996. — 307 p. 4. Adamatskii A. Artificial Life Models in Software / A. Adamatskii, M. Komosinski // Springer-Verlag. 5. Adami C. Introduction to Artificial Life / C. Adami // Springer. — Berlin, 2005. 6. Aguilar W. The Past, Present, and Future of Artificial Life, Frontiers in Robotics and AI V. 1. / W. Aguilar, B.G. Santamaria, T. Froese, C. Gershenson. — 2014. — Doi: 10.3389/frobt.2014.00008. 7. Bedau M.A. Open problems in artificial life. Artif. Life 6 / M.A. Bedau, J.S. McCaskill, N.H. Packard, S. Rasmussen, C. Adami, G. Green. — 2000. — P. 363–376. — Doi:10.1162/106454600300103683, doi:10.3389/frobt.2014. 00008. 8. Langton C.G. Artificial Life: An Overview / C.G. Langton. — Cambridge, MA: MIT Press, 1997. 9. Penn A. Artificial Life and Society: Philosophies and Tools for Experiencing, Inter- acting and Managing Real World Complex Adaptive Systems. Proc. Int. Conf. Acai / A. Penn. — Mexico: Cancun, 2016. — P. 26–27. 10. Taylor T. A Review of the First 21 Year of Artificial Life on the Web / T. Taylor, J.E. Auerbach, J. Bongard // Artificial Life (USA). — Vol. 22, N 3. — Massacussets: MIT, 2016. — P. 364–407. 11. Wikipedia. Artificial Life. — Avaliable at: https://en.wikipedia.org/ wiki/ Artificial_life 12. Dorin A. Biological Bits. A Brief Guide to the Ideas and Artefacts of Computational Artificial Life / A. Dorin. — Melbourne: Animaland, 2014. 13. Gras R. Speciation without Pre-Defined Fitness Functions. PLoS ONE 10(9): e0137838 / R. Gras, A. Golestani, A.P. Hendry, M.E. Cristescu. 2015. — Doi: 10.1371/journal. pone.0137838 14. Krivenko S. Simulation of the evolution of aging: effects of aggression and kin- recognition. In Advances in Artificial Life / S. Krivenko, M. Burtsev // 9th European Conference, ECAL, Lecture Notes in Computer Science, 2007. — P. 84–92. 15. Lindgren K. Cooperation and Community structure in artificial ecosystems / K. Lindgren, N.G. Nordahl // Artificial Life, 1994. — P. 15–37. 16. D’Ambrosio D.B. Generative encoding for multiagent learning / D.B. D’Ambrosio, K.O. Stanley // In Proceedings of the Genetic and Evolutionary Computation Conference. — New York: ACM Press, 2008. — Doi: 10.1145/1389095. 1389256. 17. Yaeger L. Computational Genetics, Physiology, Learning, Vision, and Behavior or PolyWord: Life in a New Context. In Langton, C. G. (ed.) / L. Yaeger // Artifi- cial Life III. Addison-Wesley, 1994. — P. 263–298. Genotype dynamic for agent neuroevolution in artificial life model … Системні дослідження та інформаційні технології, 2017, № 1 87 18. Packard N. Intrinsic adaptation in a simple model for evolution / N. Packard, C.G. Langton // Artificial life, Redwood City, Addison-Wesley, CA, 1989. — P. 141–155. 19. Forrest S. Modeling complex adaptive systems with Echo. Complex Systems: Mechanisms of Adaptation / S. Forrest, T. Jones. — Amsterdam: IOS Press, 1994. — P. 3–21. 20. Hraber P.T. Community Assembly in a Model Ecosystem. / P.T. Hraber, B.T. Milne // Ecological Modeling. — 1997. —P. 267–285. 21. Burtsev M. Evolution of Cooperative Strategies From First Principles / M. Burtsev, P. Turchin // Nature. 2006. — P. 1041–1044. 22. Zavertanyy V. Aggressive and peaceful behavior in multiagent systems on cellu- lar space / V. Zavertanyy, A. Makarenko // Cистемні дослідження та інформаційні технології. — № 2. — 2016. — P. 36–44. Received 31.10.2016 From the Editorial Board: the article corresponds completely to submitted manuscript.

Genotype dynamic for agent neuroevolution in artificial life model

Репозитарії

Схожі ресурси