Dr. Stefan Edelkamp

Effective decision making while competing for limited resources in adversarial environments is important for many real-world applications (e.g. two Taxi companies competing for customers). Decision-making techniques such as Automated planning have to take into account possible actions of adversary (or competing) agents. That said, the agent should know what the competitor will likely do and then generate its plan accordingly. In this paper we propose a novel approach for estimating strategies of the adversary (or the competitor), sampling its actions that might hinder agent's goals by interfering with the agent's actions. The estimated competitor strategies are used in plan generation such that agent's actions have to be applied prior to the ones of the competitor, whose estimated times dictate the deadlines. We empirically evaluate our approach leveraging sampling of competitor's actions by comparing it to the naive approach optimizing the make-span (not taking the competing agent into account at all) and to Nash Equilibrium (mixed) strategies.

Sampling-based motion planning algorithms such as Rapidly exploring Random Trees (RRTs) have been used in robotic applications for a long time. In this paper, we propose a method that combines deep learning with RRT* method. We use a neural network to learn a sample strategy for RRT*.We evaluate Deep RRT* in a collection of 2D scenarios. The results demonstrate that our algorithm could find collision-free paths efficiently and fast, and can be generalized to unseen environments.

Effective planning while competing for limited resources is crucial in many real-world applications such as on-demand transport companies competing for passengers. Planning techniques therefore have to take into account possible actions of an adversarial agent. Such a challenge that can be tackled by leveraging game-theoretical methods such as Double Oracle. This paper aims at the scalability issues arising from combining planning techniques with Double Oracle. In particular, we propose an abstraction-based heuristic for deciding how resources will be collected (e.g. which car goes for which passenger and in which order) and we propose a method for decomposing planning tasks into smaller ones (e.g. generate plans for each car separately). Our empirical evaluation shows that our proposed approach considerably improves scalability compared to the state-of-the-art techniques.

This paper shows that domain-independent tools from classical planning can be used to model and solve a broad class of game-theoretic problems we call Cost-Adversarial Planning Games (CAPGs). We define CAPGs as 2-player normal-form games specified by a planning task and a finite collection of cost functions. The first player (a planning agent) strives to solve a planning task optimally but has limited knowledge about its action costs. The second player (an adversary agent) controls the actual action costs. Even though CAPGs need not be zero-sum, every CAPG has an associated zero-sum game whose Nash equilibrium provides the optimal randomized strategy for the planning agent in the original CAPG. We show how to find the Nash equilibrium of the associated zero-sum game using a cost-optimal planner via the Double Oracle algorithm. To demonstrate the expressivity of CAPGs, we formalize a patrolling security game and several IPC domains as CAPGs.

Effective plan generation in adversarial environments has totake into account possible actions of adversary agents, i.e.,the agent should know what the competitor will likely do.In this paper we propose a novel approach for estimatingstrategies of the adversary, sampling actions that interferewith the agent’s ones. The estimated competitor strategies areused in plan generation by considering that agent’s actionshave to be applied prior to the ones of the competitor, whoseestimated times dictate the agent’s deadlines. Missing thesedeadlines entails additional plan cost