-
Wasserstein Robust Reinforcement Learning
Authors:
Mohammed Amin Abdullah,
Hang Ren,
Haitham Bou Ammar,
Vladimir Milenkovic,
Rui Luo,
Mingtian Zhang,
Jun Wang
Abstract:
Reinforcement learning algorithms, though successful, tend to over-fit to training environments hampering their application to the real-world. This paper proposes $\text{W}\text{R}^{2}\text{L}$ -- a robust reinforcement learning algorithm with significant robust performance on low and high-dimensional control tasks. Our method formalises robust reinforcement learning as a novel min-max game with a…
▽ More
Reinforcement learning algorithms, though successful, tend to over-fit to training environments hampering their application to the real-world. This paper proposes $\text{W}\text{R}^{2}\text{L}$ -- a robust reinforcement learning algorithm with significant robust performance on low and high-dimensional control tasks. Our method formalises robust reinforcement learning as a novel min-max game with a Wasserstein constraint for a correct and convergent solver. Apart from the formulation, we also propose an efficient and scalable solver following a novel zero-order optimisation method that we believe can be useful to numerical optimisation in general. We empirically demonstrate significant gains compared to standard and robust state-of-the-art algorithms on high-dimensional MuJuCo environments.
△ Less
Submitted 16 September, 2019; v1 submitted 30 July, 2019;
originally announced July 2019.
-
Geometric Rounding and Feature Separation in Meshes
Authors:
Victor Milenkovic,
Elisha Sacks
Abstract:
Geometric rounding of a mesh is the task of approximating its vertex coordinates by floating point numbers while preserving mesh structure. Geometric rounding allows algorithms of computational geometry to interface with numerical algorithms. We present a practical geometric rounding algorithm for 3D triangle meshes that preserves the topology of the mesh. The basis of the algorithm is a novel str…
▽ More
Geometric rounding of a mesh is the task of approximating its vertex coordinates by floating point numbers while preserving mesh structure. Geometric rounding allows algorithms of computational geometry to interface with numerical algorithms. We present a practical geometric rounding algorithm for 3D triangle meshes that preserves the topology of the mesh. The basis of the algorithm is a novel strategy: 1) modify the mesh to achieve a feature separation that prevents topology changes when the coordinates change by the rounding unit; and 2) round each vertex coordinate to the closest floating point number. Feature separation is also useful on its own, for example for satisfying minimum separation rules in CAD models. We demonstrate a robust, accurate implementation.
△ Less
Submitted 8 May, 2018;
originally announced May 2018.
-
Table Based Detection of Degenerate Predicates in Free Space Construction
Authors:
Victor Milenkovic,
Elisha Sacks,
Nabeel Butt
Abstract:
The key to a robust and efficient implementation of a computational geometry algorithm is an efficient algorithm for detecting degenerate predicates. We study degeneracy detection in constructing the free space of a polyhedron that rotates around a fixed axis and translates freely relative to another polyhedron. The structure of the free space is determined by the signs of univariate polynomials,…
▽ More
The key to a robust and efficient implementation of a computational geometry algorithm is an efficient algorithm for detecting degenerate predicates. We study degeneracy detection in constructing the free space of a polyhedron that rotates around a fixed axis and translates freely relative to another polyhedron. The structure of the free space is determined by the signs of univariate polynomials, called angle polynomials, whose coefficients are polynomials in the coordinates of the vertices of the polyhedra. Every predicate is expressible as the sign of an angle polynomial $f$ evaluated at a zero $t$ of an angle polynomial $g$. A predicate is degenerate (the sign is zero) when $t$ is a zero of a common factor of $f$ and $g$. We present an efficient degeneracy detection algorithm based on a one-time factoring of every possible angle polynomial. Our algorithm is 3500 times faster than the standard algorithm based on greatest common divisor computation. It reduces the share of degeneracy detection in our free space computations from 90% to 0.5% of the running time.
△ Less
Submitted 19 March, 2018;
originally announced March 2018.
-
Planar shape manipulation using approximate geometric primitives
Authors:
Victor Milenkovic,
Elisha Sacks,
Steven Trac
Abstract:
We present robust algorithms for set operations and Euclidean transformations of curved shapes in the plane using approximate geometric primitives. We use a refinement algorithm to ensure consistency. Its computational complexity is $\bigo(n\log n+k)$ for an input of size $n$ with $k=\bigo(n^2)$ consistency violations. The output is as accurate as the geometric primitives. We validate our algorith…
▽ More
We present robust algorithms for set operations and Euclidean transformations of curved shapes in the plane using approximate geometric primitives. We use a refinement algorithm to ensure consistency. Its computational complexity is $\bigo(n\log n+k)$ for an input of size $n$ with $k=\bigo(n^2)$ consistency violations. The output is as accurate as the geometric primitives. We validate our algorithms in floating point using sequences of six set operations and Euclidean transforms on shapes bounded by curves of algebraic degree~1 to~6. We test generic and degenerate inputs.
Keywords: robust computational geometry, plane subdivisions, set operations.
△ Less
Submitted 1 October, 2012;
originally announced October 2012.
