In this paper we study the strengths and limitations of collaborative teams of simple agents. In particular, we discuss the efficient use of "ant robots" for covering a connected region on the $Z^{2}$ grid, whose area is unknown in advance and which expands stochastically. Specifically, we discuss the problem where an initial connected region of $S_0$ boundary tiles expand outward with probability $p$ at every time step. On this grid region a group of $k$ limited and simple agents operate, in order to clean the unmapped and dynamically expanding region. A preliminary version of this problem was discussed in [1],[2] involving a deterministic expansion of a region in the grid.In this work we extend the model and examine cases where the spread of the region is done stochastically, where each tile has some probability $p$ to expand, at every time step. For this extended model we obtain an analytic probabilistic lower bounds for the minimal number of agents and minimal time required to enable a collaborative coverage of the expanding region, regardless of the algorithm used and the robots' hardware and software specifications. In addition, we present an impossibility result, for a variety of regions that would be impossible to completely clean, regardless of the algorithm used. Finally, we validate the analytic bounds using extensive empirical computer simulation results.
Cooperative Control of Mobile Robots with Stackelberg Learning
University of Colorado Boulder（科罗拉多大学博尔德分校）
多机器人合作要求代理做出与共同目标一致的决策，而不忽视可能因能力和个人目标的不对称而产生的特定于行动的偏好。为了实现这一目标，我们提出了一种名为SLiCC的方法：Stackelberg Learning in Cooperative Control。SLiCC将问题建模为由Stackelberg双矩阵游戏组成的部分可观察随机游戏，并使用深度强化学习来获得与这些游戏相关的支付矩阵。适当的合作行动，然后选择与派生的Stackelberg均衡。使用一个双机器人协作对象运输问题，我们验证了SLiCC对集中式多智能体Q学习的性能，并证明SLiCC实现了更好的组合效用。
Robust Instruction Compliance in Cooperative Multi-Agent Reinforcement Learning
Department of Computer Sciences, Northeastern University（东北大学计算机科学系）；Department of Computer Sciences, Massachusetts Institute of Technology（麻省理工学院计算机科学系）
