Evolutionary learning in complex social networks

复杂系统和基于智能体的建模正吸引着来自不同学科的专家对模拟和研究现实世界的应用越来越感兴趣。这些影响具有重要意义，因为它们拓宽了我们对鲜为人知的复杂社会现象的理解，包括对观察到的系统行为的原因进行推理，以及预测测试案例情景或干预的结果。这一领域包括的令人兴奋的系统包括：恐怖分子网络、史前文明和车辆安全。例如，代理人可以代表个人、家庭或司机。系统设计人员经常为代理构建基于状态的配置，以便在动态环境中做出反应。因此，该模型为研究人员检验各种假设提供了一种沙盒方法。例如，可以使用真实的实地数据来模拟古代文明，考古学家随后可以检查干旱条件下与正常条件相比对新出现的人口数量和迁移的影响，以检验关于干旱导致人口离开研究区域的假设。该研究计划建议在复杂系统中发展代理行为时，将新技术应用于代理学习。除了代理可以用来对其动态环境做出反应的基本有限状态机之外，更现实的代理智能方法提出了一种反应式代理，它可以在个人和种群级别积累知识，学习新的计划，并自行进化新的策略来实现自己的目标。然而，人类行为与其他动物行为不同，因此需要扩展传统的进化框架，以建立更现实的人工代理人。通过所提出的方法，本研究项目的成果将加强人工模型中的学习理论，研究人员将能够使用本工作中的算法创建更准确和更智能的模型。

Complex systems and agent-based modelling are attracting a growing interest from experts from various disciplines to simulate and study real world applications. The implications are of significance because they broaden our understanding of little known complex social phenomena, including reasoning about the causes of observed system behaviour and predicting the outcomes of tested case scenarios or interventions. Exciting systems that this field encompasses include: terrorist networks, prehistoric civilizations, and vehicle safety. An agent for instance can represent an individual, a household, or a driver. System designers often build a state based configuration for agents to react in a dynamic environment. Hence, the model provides a sandbox approach for the researchers to test various hypotheses. For example, an ancient civilization can be modelled using real field data where an archaeologist can then examine the effects on the emerging population size and movement under drought conditions compared to normal conditions in order to test a hypothesis suggesting that a drought caused the population to leave the study area. The research program proposes applying new techniques in agent learning when developing agent behaviour in a complex system. Beyond the basic finite state machine that an agent can use to react to its dynamic environment, a more realistic approach for agent intelligence proposes reactive agents that can accumulate knowledge at the individual and population levels, learn new plans and evolve new strategies on their own to fulfill their goals. Human behaviour however, is distinguished from other animal behaviours and consequently the conventional evolutionary frameworks need to be extended in order to build more realistic artificial agents. Through the proposed methods, the outcome of this research program will enhance the learning theory in artificial models and researchers working with agent-based simulations would be able to create more accurate and intelligent models using the algorithms from this work.