S&AS: FND: Decision-Making for Autonomous Systems with Limited Resources

自主机器人系统决策中的一个基本问题是，这些系统如何做出决策，使这些决策与人类在类似情况下做出的决策相似。决策的这一方面对于机器人代理产生可预测的行为是重要的，这些行为是建立对人机团队的信任所必需的。解决这个问题的传统方法假设人类是消息灵通的、完全理性的代理人，总是达到最优(例如，最佳决策)。然而，一些实证研究表明，大多数人在做出决策时并不是完全理性的代理人，而是有限理性的代理人。换句话说，他们不试图无条件地获得可能的最佳结果，而是倾向于权衡在有限制(时间、认知努力、精力、金钱或只是懒惰)的情况下实现最佳结果的好处。基于这一核心思想，本研究将研究资源约束下自治系统的决策问题。与人类做出决策的方式类似，决策过程将必须考虑代理可用的计算、信息或能源限制，而不是无论如何都要试图获得最优决策。本项目研究自主机器人智能体与其他人类或机器人智能体之间的交互，同时考虑智能体的决策约束。这项研究不是关注预期效用最大化(如在标准公式中那样)，而是关注预期自由效用(或自由能量)的最大化，这是一个以精确方式封装决策过程的信息和计算限制的影响的量。这项研究的结果将应用于多个代理可能具有相互冲突的目标的情况。作为测试用例，该框架被应用于交通中的自动驾驶车辆问题，包括人类驾驶员和其他自动驾驶车辆。通过使代理的决策过程适应可用的资源，该框架可用于生成信息处理成本的严重性所带来的决策的抽象层次结构。每个抽象都针对问题的“正确”粒度级别进行了调整。也就是说，当计算资源稀缺时，首选较粗的抽象级别，而当计算资源充足时，首选较细(更准确)的抽象级别。最后，“心理理论”原则将被用来设计有限理性的决策算法，以避免冲突主体之间的无限回归。这种无限的倒退阻碍了纯粹理性均衡的计算，并将通过采用限制每个参与者递归推理深度的框架来避免。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

A fundamental question in decision-making of autonomous robotic systems is how these systems make decisions such that these are similar to a decision a human would make in a similar situation. This aspect of decision making is important in order for the robotic agent to generate predictable behaviors which are necessary in order to establish trust in human-machine teams. Traditional approaches to solve this problem assume that humans are well-informed, perfectly rational agents that always reach the optimal (e.g., the best decision). Several empirical studies, however, suggest that most humans do not act as perfectly rational agents when making decisions but rather as bounded-rational ones. In other words, they do not try to get the best possible outcome unconditionally, but rather they tend to weight the benefits of reaching the best outcome subject to constraints (time, cognitive effort, energy, money, or just laziness). Using this key idea, this research will investigate decision-making for autonomous systems subject to resource constraints. Similar to how humans make decisions, instead of trying to get the optimal decision no matter what, the decision-making process will have to factor in the computational, information, or energy constraints available to the agent. This project investigates the interaction between autonomous robotic agents and other agents either human or robotic, while considering the decision-making constraints of the agents. Instead of focusing on maximizing expected utility (as in standard formulations) this research will address the maximization of expected free utility (or free energy), a quantity that encapsulates in a precise manner the effect of information and computational constraints of the decision-making process. The results of this research will be applied to the case of multiple agents having perhaps conflicting objectives. As a test case this framework is applied to the problem of autonomous vehicles in traffic including both human drivers and other self-driving vehicles. By adapting the agent's decision-making process to the available resources, this framework can be used to generate a hierarchy of abstractions for decision-making brought about by the severity of information-processing costs. Each abstraction is tailored to the "right" level of granularity of the problem. That is, coarser abstraction levels are preferred when the computational resources are scarce and finer (more accurate) levels of abstraction when the computational resources are plentiful. Finally, "theory of mind" principles will be utilized to design bounded-rational decision-making algorithms that avoid the infinite regress between conflicting agents. This infinite regress hinders the computation of purely rational equilibria and will be avoided by adopting a framework that restricts each player's depth of recursive inference.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.