Collective behaviour of cognitive agents

动物的行为是由进化适应驱动的，通过食物获取，配偶选择和捕食者回避等目标，使适应性最大化。在许多物种中，认知在让动物感知世界、处理信息和计划行动方面起着关键作用。该研究金将根据理性决策、贝叶斯推理和人工智能的模型，发展一种关于代理人对来自其物理和社会环境的信息作出反应的认知驱动行为的综合理论。最近的技术发展大大增加了动物行为研究人员可用的数据。研究人员正在利用这些数据来了解动物如何与环境以及彼此相互作用。然而，一种机械的、数据驱动的方法，专注于从直接刺激中预测行为，忽视了动物学习世界、制定策略和执行行为模式以达到目标的认知能力。这反过来又会导致关于行为的误导性结论，这些结论并没有超出研究它们的特定实验场景。只有发展出一种明确包括认知过程和动物试图实现的最终目标的行为理论，我们才能理解为什么动物会以我们观察到的方式行为，为什么它们对某些刺激的反应或多或少强烈，以及当动物面对不同的环境时，它们的行为会发生怎样的变化。这种理论必须建立在原则和观察的基础上。在这个奖学金中，我将建立假设代理是理性的适应度最大化者，能够对世界进行推理，并根据经验合理地更新他们的信念，并能够通过使用对其行为可能后果的预测来做出战略决策。我将模拟智能体如何在不确定的世界中运作的同时，做出优化长期目标（如健身最大化）的决策。在这个框架中，代理人必须根据他们对世界的了解来考虑他们的行为的后果。即时回报较低的行为可能会在以后通过获得的知识开辟新的可能性。例如，觅食动物可能会选择探索新的领域，而不是利用已知的资源，希望找到更有生产力的地区。由于所有行动的结果都是不确定的，因此智能体必须权衡所有可能的结果可以获得和学习什么，并决定如何平衡这些可能性以优化其长期目标。再加上这个模型的个人战略行为，我也将开发代理人的互动，合作和竞争的模型。在这里，代理人必须在追求自己目标的其他人的背景下追求自己的目标。这些人可能有相似或不同的特点，他们的目标可能是一致的或不同的。在每种情况下，代理人采取的最佳行动必须基于其他人采取的行动的模型，这传达了什么信息，以及他们将来可能采取什么行动。这个理论框架将揭示为了验证或证伪理论所基于的基本假设所需的观察结果。因此，该奖学金将通过确定实验设计的原则来推动下一代实证研究，以最大限度地提高数据的有用性，并走向精确目标数据收集的时代。成功的经验验证的模式也将允许他们用于预测的结果，外部变化的变化，个人和群体的行为。

Animal behaviour is driven by evolutionary adaptations that maximise fitness, through goals such as food acquisition, mate selection and predator avoidance. In many species cognition plays a key role in allowing animals to perceive the world, process information and plan actions. This fellowship will develop a comprehensive theory of cognitively-driven behaviour by agents responding to information from their physical and social environment, based on models of rational decision-making, Bayesian inference and artificial intelligence. Recent technological developments have dramatically increased the data available to researchers of animal behaviour. Researchers are using these data to understand how animals interact with their environment and with each other. However, a mechanistic, data-driven approach that focuses on predicting behaviour from the immediate stimuli ignores the cognitive ability of animals to learn about their world, strategise and execute patterns of behaviour to reach their goals. This in turn can lead to misleading conclusions about behaviour that do not generalise beyond the specific experimental scenario under which they were studied.Only by developing a theory of behaviour that explicitly includes cognitive processes and the end goals that animals are trying to accomplish can we understand why animals behave in the ways that we observe, why they respond more or less strongly to certain stimuli, and how behaviour may change when animals are confronted by a different environment. Such a theory must build on principles as well as observations. In this fellowship I will establish assume that agents are rational fitness-maximisers, capable of performing inference about the world and rationally updating their beliefs on the basis of experience, and able to make strategic decisions through the use of predictions about the likely consequences of their actions.Building on developments in artificial intelligence known as active-learning, I will model how an agent takes decisions to optimise long-term goals, such as fitness maximisation, while operating in a world of uncertainty. In this framework, agents must consider the consequences of their actions in terms of what they learn about the world. Actions with lower immediate reward may open up new possibilities later through knowledge gained. For example, a foraging animal may choose to explore new territory rather than exploit a known resource, hoping to find more productive areas. Because the outcomes of all actions are uncertain, an agent must weigh up what can be gained and learnt for all possible results, and decide how to balance these possibilities to optimise its long-term objectives. Coupled with this model of individual strategic behaviour, I will also develop models of agents interacting, cooperating and competing. Here, agents must pursue their individual objectives in the context of others who are pursuing their own goals. These individuals may have similar or differing characteristics, and their goals may be aligned or divergent. In each case, the best action for an agent to take must be based on a model of the actions others have taken, what information this conveys, and what actions they are likely to take in the future. This theoretical framework will reveal the observations that are required in order to verify or falsify the underlying assumptions on which the theory is built. Thus this fellowship will drive the next generation of empirical studies by identifying principles of experimental design to maximise the usefulness of data, and move towards an era of precision-targeted data collection. Successful empirical validation of the models will also allow their use for predicting the outcome of exogenous changes by in terms of changes to individual and group behaviours.