A key challenge in multi-agent systems is the design of intelligent agents solving real-world tasks in close interaction with other agents (e.g. humans), thereby being confronted with a variety of behavioral variations and limited knowledge about the true behaviors of observed agents. The practicability of existing works addressing this challenge is being limited due to using finite sets of hypothesis for behavior prediction, the lack of a hypothesis design process ensuring coverage over all behavioral variations and sample-inefficiency when modeling continuous behavioral variations. In this work, we present an approach to this challenge based on a new framework of Robust Stochastic Bayesian Games (RSBGs). An RSBG defines hypothesis sets by partitioning the physically feasible, continuous behavior space of the other agents. It combines the optimality criteria of the Robust Markov Decision Process (RMDP) and the Stochastic Bayesian Game (SBG) to exponentially reduce the sample complexity for planning with hypothesis sets defined over continuous behavior spaces. Our approach outperforms the baseline algorithms in two experiments modeling time-varying intents and large multidimensional behavior spaces, while achieving the same performance as a planner with knowledge of the true behaviors of other agents.

中文翻译：

---

行为空间覆盖的鲁棒随机贝叶斯博弈

多代理系统中的一个关键挑战是智能代理的设计，以与其他代理（例如人类）密切交互来解决现实世界的任务，从而面临各种行为变化和对观察到的代理真实行为的有限了解。由于使用有限的假设集进行行为预测，缺乏确保覆盖所有行为变化的假设设计过程，以及在对连续行为变化建模时样本效率低下，解决这一挑战的现有工作的实用性受到限制。在这项工作中，我们提出了一种基于鲁棒随机贝叶斯博弈 (RSBG) 新框架来应对这一挑战的方法。RSBG 通过划分其他代理的物理上可行的连续行为空间来定义假设集。它结合了鲁棒马尔可夫决策过程 (RMDP) 和随机贝叶斯博弈 (SBG) 的最优性标准，以指数方式降低规划的样本复杂度，假设集定义在连续行为空间上。我们的方法在两个模拟时变意图和大型多维行为空间的实验中优于基线算法，同时实现了与了解其他代理真实行为的规划者相同的性能。