-
Grouping Shapley Value Feature Importances of Random Forests for explainable Yield Prediction
Authors:
Florian Huber,
Hannes Engler,
Anna Kicherer,
Katja Herzog,
Reinhard Töpfer,
Volker Steinhage
Abstract:
Explainability in yield prediction helps us fully explore the potential of machine learning models that are already able to achieve high accuracy for a variety of yield prediction scenarios. The data included for the prediction of yields are intricate and the models are often difficult to understand. However, understanding the models can be simplified by using natural groupings of the input featur…
▽ More
Explainability in yield prediction helps us fully explore the potential of machine learning models that are already able to achieve high accuracy for a variety of yield prediction scenarios. The data included for the prediction of yields are intricate and the models are often difficult to understand. However, understanding the models can be simplified by using natural groupings of the input features. Grouping can be achieved, for example, by the time the features are captured or by the sensor used to do so. The state-of-the-art for interpreting machine learning models is currently defined by the game-theoretic approach of Shapley values. To handle groups of features, the calculated Shapley values are typically added together, ignoring the theoretical limitations of this approach. We explain the concept of Shapley values directly computed for predefined groups of features and introduce an algorithm to compute them efficiently on tree structures. We provide a blueprint for designing swarm plots that combine many local explanations for global understanding. Extensive evaluation of two different yield prediction problems shows the worth of our approach and demonstrates how we can enable a better understanding of yield prediction models in the future, ultimately leading to mutual enrichment of research and application.
△ Less
Submitted 14 April, 2023;
originally announced April 2023.
-
On the Lorenz '96 Model and Some Generalizations
Authors:
John Kerin,
Hans Engler
Abstract:
In 1996, Edward Lorenz introduced a system of ordinary differential equations that describes a single scalar quantity as it evolves on a circular array of sites, undergoing forcing, dissipation, and rotation invariant advection. Lorenz constructed the system as a test problem for numerical weather prediction. Since then, the system has also found widespread use as a test case in data assimilation.…
▽ More
In 1996, Edward Lorenz introduced a system of ordinary differential equations that describes a single scalar quantity as it evolves on a circular array of sites, undergoing forcing, dissipation, and rotation invariant advection. Lorenz constructed the system as a test problem for numerical weather prediction. Since then, the system has also found widespread use as a test case in data assimilation. Mathematically, it belongs to a class of dynamical systems with a single bifurcation parameter (rescaled forcing) that undergoes multiple bifurcations and exhibits chaotic behavior for large forcing. In this paper, the main characteristics of the advection term in the model are identified and used to describe and classify a number of possible generalizations of the system. A graphical method to study the bifurcation behavior of constant solutions is introduced, and it is shown how to use the rotation invariance to compute normal forms of the system analytically. Problems with site-dependent forcing, dissipation, or advection are considered and basic existence and stability results are proved for these extensions. We address some related topics in the appendices, wherein the Lorenz '96 system in Fourier space is considered, explicit solutions for some advection-only systems are found, and it is demonstrated how to use advection-only systems to assess numerical schemes.
△ Less
Submitted 21 October, 2020; v1 submitted 15 May, 2020;
originally announced May 2020.
-
Modeling the Dynamics of Glacial Cycles
Authors:
Hans Engler,
Hans G. Kaper,
Tasso J. Kaper,
Theodore Vo
Abstract:
This article is concerned with the dynamics of glacial cycles observed in the geological record of the Pleistocene Epoch. It focuses on a conceptual model proposed by Maasch and Saltzman [J. Geophys. Res.,95, D2 (1990), pp. 1955-1963], which is based on physical arguments and emphasizes the role of atmospheric CO2 in the generation and persistence of periodic orbits (limit cycles). The model consi…
▽ More
This article is concerned with the dynamics of glacial cycles observed in the geological record of the Pleistocene Epoch. It focuses on a conceptual model proposed by Maasch and Saltzman [J. Geophys. Res.,95, D2 (1990), pp. 1955-1963], which is based on physical arguments and emphasizes the role of atmospheric CO2 in the generation and persistence of periodic orbits (limit cycles). The model consists of three ordinary differential equations with four parameters for the anomalies of the total global ice mass, the atmospheric CO2 concentration, and the volume of the North Atlantic Deep Water (NADW). In this article, it is shown that a simplified two-dimensional symmetric version displays many of the essential features of the full model, including equilibrium states, limit cycles, their basic bifurcations, and a Bogdanov-Takens point that serves as an organizing center for the local and global dynamics. Also, symmetry breaking splits the Bogdanov-Takens point into two, with different local dynamics in their neighborhoods.
△ Less
Submitted 21 May, 2017;
originally announced May 2017.
-
Dynamical systems analysis of the Maasch-Saltzman model for glacial cycles
Authors:
Hans Engler,
Hans G. Kaper,
Tasso J. Kaper,
Theodore Vo
Abstract:
This article is concerned with the internal dynamics of a conceptual model proposed by Maasch and Saltzman [J. Geophys. Res., 95, D2 (1990) 1955-1963] to explain central features of the glacial cycles observed in the climate record of the Pleistocene Epoch. It is shown that, in most parameter regimes, the long-term system dynamics occur on certain intrinsic two-dimensional invariant manifolds in t…
▽ More
This article is concerned with the internal dynamics of a conceptual model proposed by Maasch and Saltzman [J. Geophys. Res., 95, D2 (1990) 1955-1963] to explain central features of the glacial cycles observed in the climate record of the Pleistocene Epoch. It is shown that, in most parameter regimes, the long-term system dynamics occur on certain intrinsic two-dimensional invariant manifolds in the three-dimensional state space. These invariant manifolds are slow manifolds when the characteristic time scales for the total global ice mass and the volume of North Atlantic Deep Water are well- separated, and they are center manifolds when the characteristic time scales for the total global ice mass and the volume of North Atlantic Deep Water are comparable. In both cases, the reduced dynamics on these manifolds are governed by Bogdanov-Takens singularities, and the bifurcation curves associated to these singularities organize the parameter regions in which the model exhibits glacial cycles.
This work was submitted March 30, 2017.
△ Less
Submitted 17 May, 2017;
originally announced May 2017.
-
Computation of Scattering Kernels in Radiative Transfer
Authors:
Hans Engler
Abstract:
This note proposes rapidly convergent computational formulae for evaluating scattering kernels from radiative transfer theory. The approach used here does not rely on Legendre expansions, but rather uses exponentially convergent numerical integration rules. A closed form for the Henyey-Greenstein scattering kernel in terms of complete elliptic integrals is also derived.
This note proposes rapidly convergent computational formulae for evaluating scattering kernels from radiative transfer theory. The approach used here does not rely on Legendre expansions, but rather uses exponentially convergent numerical integration rules. A closed form for the Henyey-Greenstein scattering kernel in terms of complete elliptic integrals is also derived.
△ Less
Submitted 10 January, 2015;
originally announced January 2015.
-
A characterization of the behavior of the Anderson acceleration on linear problems
Authors:
Florian Potra,
Hans Engler
Abstract:
We give a complete characterization of the behavior of the Anderson acceleration (with arbitrary nonzero mixing parameters) on linear problems. Let n be the grade of the residual at the starting point with respect to the matrix defining the linear problem. We show that if Anderson acceleration does not stagnate (that is, produces different iterates) up to n, then the sequence of its iterates conve…
▽ More
We give a complete characterization of the behavior of the Anderson acceleration (with arbitrary nonzero mixing parameters) on linear problems. Let n be the grade of the residual at the starting point with respect to the matrix defining the linear problem. We show that if Anderson acceleration does not stagnate (that is, produces different iterates) up to n, then the sequence of its iterates converges to the exact solution of the linear problem. Otherwise, the Anderson acceleration converges to the wrong solution. Anderson acceleration and of GMRES are essentially equivalent up to the index where the iterates of Anderson acceleration begin to stagnate. This result holds also for an optimized version of Anderson acceleration, where at each step the mixing parameter is chosen so that it minimizes the residual of the current iterate.
△ Less
Submitted 3 February, 2011;
originally announced February 2011.
-
On the Speed of Spread for Fractional Reaction-Diffusion Equations
Authors:
Hans Engler
Abstract:
The fractional reaction diffusion equation u_t + Au = g(u) is discussed, where A is a fractional differential operator on the real line with order αbetween 0 and 2, the C^1 function g vanishes at 0 and 1, and either g is non-negative on (0,1) or g < 0 near 0. In the case of non-negative g, it is shown that solutions with initial support on the positive half axis spread into the left half axis wi…
▽ More
The fractional reaction diffusion equation u_t + Au = g(u) is discussed, where A is a fractional differential operator on the real line with order αbetween 0 and 2, the C^1 function g vanishes at 0 and 1, and either g is non-negative on (0,1) or g < 0 near 0. In the case of non-negative g, it is shown that solutions with initial support on the positive half axis spread into the left half axis with unbounded speed if g satisfies some weak growth condition near 0 in the case α> 1, or if g is merely positive on a sufficiently large interval near 1 in the case α< 1. On the other hand, it shown that solutions spread with finite speed if g'(0) < 0. The proofs use comparison arguments and a new family of traveling wave solutions for this class of problems.
△ Less
Submitted 31 July, 2009;
originally announced August 2009.
-
Random Search Algorithms for the Sparse Null Vector Problem
Authors:
Hans Engler
Abstract:
We consider the following problem: Given a matrix A, find minimal subsets of columns of A with cardinality no larger than a given bound that are linear dependent or nearly so. This problem arises in various forms in optimization, electrical engineering, and statistics. In its full generality, the problem is known to be NP-complete. We present a Monte Carlo method that finds such subsets with hig…
▽ More
We consider the following problem: Given a matrix A, find minimal subsets of columns of A with cardinality no larger than a given bound that are linear dependent or nearly so. This problem arises in various forms in optimization, electrical engineering, and statistics. In its full generality, the problem is known to be NP-complete. We present a Monte Carlo method that finds such subsets with high confidence. We also give a deterministic method that is capable of proving that no subsets of linearly dependent columns up to a certain cardinality exist. The performance of both methods is analyzed and illustrated with numerical experiments.
△ Less
Submitted 21 April, 2008;
originally announced April 2008.
-
On the Long Time Behavior of Second Order Differential Equations with Asymptotically Small Dissipation
Authors:
Alexandre Cabot,
Hans Engler,
Sebastien Gadat
Abstract:
We investigate the time-asymptotic properties of solutions of the differential equation x''(t) + a(t)x'(t) + g(x(t)) = 0 in a Hilbert space, where a(.) is non-increasing and g is the gradient of a potential G. If the coefficient a(.) is constant and positive, we recover the so-called ``Heavy Ball with Friction'' system. On the other hand, when a(t)=1/(t+1) we obtain the trajectories associated t…
▽ More
We investigate the time-asymptotic properties of solutions of the differential equation x''(t) + a(t)x'(t) + g(x(t)) = 0 in a Hilbert space, where a(.) is non-increasing and g is the gradient of a potential G. If the coefficient a(.) is constant and positive, we recover the so-called ``Heavy Ball with Friction'' system. On the other hand, when a(t)=1/(t+1) we obtain the trajectories associated to some averaged gradient system.
Our analysis is mainly based on the existence of some suitable energy function. When the potential G is convex and the coeffient a is non-integrable at infinity, the energy function converges to its minimum. A more stringent condition is required to obtain the convergence of the trajectories of toward some minimum point of the potential. In the one-dimensional setting, a precise description of the convergence of solutions is given for a general coercive non-convex potentials with many local minima and maxima. We show that in this case the set of initial conditions for which solutions converge to a local minimum is open and dense.
△ Less
Submitted 4 October, 2007;
originally announced October 2007.
-
Asymptotic Self-Similarity for Solutions of Partial Integrodifferential Equations
Authors:
Hans Engler
Abstract:
The question is studied whether weak solutions of linear partial integrodifferential equations approach a constant spatial profile after rescaling, as time goes to infinity. The possible limits and corresponding scaling functions are identified and are shown to actually occur. The limiting equations are fractional diffusion equations which are known to have self-similar fundamental solutions. Fo…
▽ More
The question is studied whether weak solutions of linear partial integrodifferential equations approach a constant spatial profile after rescaling, as time goes to infinity. The possible limits and corresponding scaling functions are identified and are shown to actually occur. The limiting equations are fractional diffusion equations which are known to have self-similar fundamental solutions. For an important special case, is is shown that the asymptotic profile is Gaussian and convergence holds in $L^2$, that is, solutions behave like fundamental solutions of the heat equation to leading order. Systems of integrodifferential equations occurring in viscoelasticity are also discussed, and their solutions are shown to behave like fundamental solutions of a related Stokes system. The main assumption is that the integral kernel in the equation is regularly varying in the sense of Karamata.
△ Less
Submitted 10 October, 2005;
originally announced October 2005.
-
Analysis of a model for the dynamics of prions II
Authors:
Hans Engler,
Jan Pruess,
Glenn F. Webb
Abstract:
A new mathematical model for the dynamics of prion proliferation involving an ordinary differential equation coupled with a partial integro-differential equation is analyzed, continuing earlier work. We show the well-posedness of this problem in a natural phase space, i.e. there is a unique global semiflow in the phase space associated to the problem.
A theorem of threshold type is derived for…
▽ More
A new mathematical model for the dynamics of prion proliferation involving an ordinary differential equation coupled with a partial integro-differential equation is analyzed, continuing earlier work. We show the well-posedness of this problem in a natural phase space, i.e. there is a unique global semiflow in the phase space associated to the problem.
A theorem of threshold type is derived for this model which is typical for mathematical epidemics. If a certain combination of kinetic parameters is below or at the threshold, there is a unique steady state, the disease-free equilibrium, which is globally asymptotically stable; above the threshold it is unstable, and there is another unique steady state, the disease equilibrium, which inherits that property.
△ Less
Submitted 27 July, 2005;
originally announced July 2005.
-
Asymptotic stability of traveling wave solutions for perturbations with algebraic decay
Authors:
Hans Engler
Abstract:
For a class of scalar partial differential equations that incorporate convection, diffusion, and possibly dispersion in one space and one time dimension, the stability of traveling wave solutions is investigated. If the initial perturbation of the traveling wave profile decays at an algebraic rate, then the solution is shown to converge to a shifted wave profile at a corresponding temporal algeb…
▽ More
For a class of scalar partial differential equations that incorporate convection, diffusion, and possibly dispersion in one space and one time dimension, the stability of traveling wave solutions is investigated. If the initial perturbation of the traveling wave profile decays at an algebraic rate, then the solution is shown to converge to a shifted wave profile at a corresponding temporal algebraic rate, and optimal intermediate results that combine temporal and spatial decay are obtained. The proofs are based on a general interpolation principle which says that algebraic decay results of this form always follow if exponential temporal decay holds for perturbation with exponential spatial decay and the wave profile is stable for general perturbations.
△ Less
Submitted 2 March, 2001;
originally announced March 2001.
-
Very long storage times and evaporative cooling of cesium atoms in a quasi-electrostatic dipole trap
Authors:
H. Engler,
T. Weber,
M. Mudrich,
R. Grimm,
M. Weidemueller
Abstract:
We have trapped cesium atoms over many minutes in the focus of a CO$_2$-laser beam employing an extremely simple laser system. Collisional properties of the unpolarized atoms in their electronic ground state are investigated. Inelastic binary collisions changing the hyperfine state lead to trap loss which is quantitatively analyzed. Elastic collisions result in evaporative cooling of the trapped…
▽ More
We have trapped cesium atoms over many minutes in the focus of a CO$_2$-laser beam employing an extremely simple laser system. Collisional properties of the unpolarized atoms in their electronic ground state are investigated. Inelastic binary collisions changing the hyperfine state lead to trap loss which is quantitatively analyzed. Elastic collisions result in evaporative cooling of the trapped gas from 25 $μ$K to 10 $μ$K over a time scale of about 150 s.
△ Less
Submitted 26 March, 2000;
originally announced March 2000.
-
Cold inelastic collisions between lithium and cesium in a two-species magneto-optical trap
Authors:
U. Schlöder,
H. Engler,
U. Schünemann,
R. Grimm,
M. Weidemüller
Abstract:
We investigate collisional properties of lithium and cesium which are simultaneously confined in a combined magneto-optical trap. Trap-loss collisions between the two species are comprehensively studied. Different inelastic collision channels are identified, and inter-species rate coefficients as well as cross sections are determined. It is found that loss rates are independent of the optical ex…
▽ More
We investigate collisional properties of lithium and cesium which are simultaneously confined in a combined magneto-optical trap. Trap-loss collisions between the two species are comprehensively studied. Different inelastic collision channels are identified, and inter-species rate coefficients as well as cross sections are determined. It is found that loss rates are independent of the optical excitation of Li, as a consequence of the repulsive Li$^*$-Cs interaction. Li and Cs loss by inelastic inter-species collisions can completely be attributed to processes involving optically excited cesium (fine-structure changing collisions and radiative escape). By lowering the trap depth for Li, an additional loss channel of Li is observed which results from ground-state Li-Cs collisions changing the hyperfine state of cesium.
△ Less
Submitted 20 February, 1999;
originally announced February 1999.
