Recent advances in deep reinforcement learning (RL) have led to considerable progress in many 2-player zero-sum games, such as Go, Poker and Starcraft. The purely adversarial nature of such games allows for conceptually simple and principled application of RL methods. However real-world settings are many-agent, and agent interactions are complex mixtures of common-interest and competitive aspects. We consider Diplomacy, a 7-player board game designed to accentuate dilemmas resulting from many-agent interactions. It also features a large combinatorial action space and simultaneous moves, which are challenging for RL algorithms. We propose a simple yet effective approximate best response operator, designed to handle large combinatorial action spaces and simultaneous moves. We also introduce a family of policy iteration methods that approximate fictitious play. With these methods, we successfully apply RL to Diplomacy: we show that our agents convincingly outperform the previous state-of-the-art, and game theoretic equilibrium analysis shows that the new process yields consistent improvements.

We designed, developed, deployed, and evaluated the Collaborative nanoManipulator (CnM), a system supporting remote collaboration between users of the nanoManipulator interface to atomic force microscopes. To be accepted by users, the shared nanoManipulator application had to have the same high level of interactivity as the single user system and the application had to support a user's ability to interleave working privately and working collaboratively. The paper describes the entire collaboration system, but focuses on the shared nanoManipulator application. Based on our experience developing the CnM, we present: a method of analyzing applications to characterize the requirements for sharing data between collaborating sites, examples of data structures that support collaboration, and guidelines for selecting appropriate synchronization and concurrency control schemes.

The Virtual-Reality Peripheral Network (VRPN) system provides a device-independent and network-transparent interface to virtual-reality peripherals. VRPN's application of factoring by function and of layering in the context of devices produces an interface that is novel and powerful. VRPN also integrates a wide range of known advanced techniques into a publicly-available system. These techniques benefit both direct VRPN users and those who implement other applications that make use of VR peripherals.

We present a system for rendering very complex 3D models at interactive rates.We select a subset of the model as preferred viewpoints and partition the space into virtual cells.Each cell contains near geometry, rendered using levels of detail and visibility culling, and far geometry, rendered as a textured depth mesh.Our system automatically balances the screen-space errors resulting from geometric simplification with those from textureddepth-mesh distortion.We describe our prefetching and data management schemes, both crucial for models significantly larger than available system memory.We have successfully used our system to accelerate walkthroughs of a 13 million triangle model of a large coal-fired power plant and of a 1.7 million triangle architectural model.We demonstrate the walkthrough of a 1.3 GB power plant model with a 140 MB cache footprint.

: Many applications in computer graphics and virtual environments need to render datasets with large numbers of primitives and high depth complexityatinteractive rates. However, standard techniques like view frustum culling and a hardware z-bu#er are unable to display datasets composed of hundred of thousands of polygons at interactive frame rates on current high-end graphics systems. We add a #conservative" visibility culling stage to the rendering pipeline, attempting to identify and avoid processing of occluded polygons. Given a moving viewpoint, the algorithm dynamically chooses a set of occluders. Each occluder is used to compute a shadow frustum, and all primitives contained within this frustum are culled. The algorithm hierarchicallytraverses the model, culling out parts not visible from the current viewpoint using e#cient, robust, and in some cases specialized interference detection algorithms. The algorithm's performance varies with the location of the viewpoint and the depth...