American Institute of Mathematical Sciences

December  2020, 7(2): 511-536. doi: 10.3934/jcd.2020021

Manifold learning for accelerating coarse-grained optimization

 1 Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08544, USA 2 Department of Applied Mathematics and Statistics (AMS), Johns Hopkins University, Baltimore, MD 21218, USA 3 Department of Mathematics, Imperial College London, London SW7 2AZ, UK 4 Department of Chemical and Biomolecular Engineering & AMS, Johns Hopkins University, Baltimore, MD 21218, USA

* Corresponding author

Received  October 2019 Published  July 2020

Algorithms proposed for solving high-dimensional optimization problems with no derivative information frequently encounter the "curse of dimensionality, " becoming ineffective as the dimension of the parameter space grows. One feature of a subclass of such problems that are effectively low-dimensional is that only a few parameters (or combinations thereof) are important for the optimization and must be explored in detail. Knowing these parameters/combinations in advance would greatly simplify the problem and its solution. We propose the data-driven construction of an effective (coarse-grained, "trend") optimizer, based on data obtained from ensembles of brief simulation bursts with an "inner" optimization algorithm, that has the potential to accelerate the exploration of the parameter space. The trajectories of this "effective optimizer" quickly become attracted onto a slow manifold parameterized by the few relevant parameter combinations. We obtain the parameterization of this low-dimensional, effective optimization manifold on the fly using data mining/manifold learning techniques on the results of simulation (inner optimizer iteration) burst ensembles and exploit it locally to "jump" forward along this manifold. As a result, we can bias the exploration of the parameter space towards the few, important directions and, through this "wrapper algorithm, " speed up the convergence of traditional optimization algorithms.

Citation: Dmitry Pozharskiy, Noah J. Wichrowski, Andrew B. Duncan, Grigorios A. Pavliotis, Ioannis G. Kevrekidis. Manifold learning for accelerating coarse-grained optimization. Journal of Computational Dynamics, 2020, 7 (2) : 511-536. doi: 10.3934/jcd.2020021
References:

show all references

References:
Evolution in time of the probability density of current points using either the Langevin equation or SA at constant $T$. The objective function is $f(x) = 0.5x^2$; $10^4$ realizations are used with starting point $x = 1$, $T = 0.5$; and the time step is $dt = 10^{-3}$
Applying Diffusion Maps to the Swiss roll data set
One complete run of the algorithm that approaches the global maximum. Six total "coarse iterations" (shown in red) were performed. In addition, a single run of SA using the same number of function evaluations is shown in green. Our algorithm visibly approaches the maximum much faster
A comparison of the evolving maximum objective value for both methods. SA combined with DMaps needs only a fraction of the total function evaluations compared to simple SA
Diffusion Maps applied to bursts on a rectangular strip]{Diffusion maps applied to (A) the original 2D data with Euclidean distances, and (B) the transformed data with Mahalanobis distances. In both cases, the inherent dimensionality is recovered in the first two eigenvectors, as the coloring of the data by the leading Diffusion Map coordinates corresponding to these two eigenvectors show
Estimation of the first drift coefficient $\theta_1$. "Theoretical" are obtained numerically from (5), "Euclidean" via DMaps on the original data, and "Mahalanobis" from DMaps on the transformed data. The last subplot shows results from Euclidean DMaps on transformed data, which, as expected, yields incorrect estimates
">Figure 7.  Estimation of the second drift coefficient $\theta_2$. The subplots are analogous to those in Figure 6
Coarse-grained (two-dimensional) optimization on a cylindrical surface in three dimensions. One complete run of the algorithm, illustrating the short bursts of trajectories and new points after being lifted by geometric harmonics. Observe that, although the lifting does not perfectly locate the cylinder, the burst trajectories are quickly attracted back to it. The second plot is a top-down view of the first
Simulation of a short trajectory initialized at perturbed initial conditions $(x_{ic}, \bf{y}_{ic})$
The first diffusion coordinate parameterizes $x$
Estimation of the drift coefficient in diffusion map space
Estimation of the diffusion coefficient in diffusion map space
Nonlinear transformation of the slow variable $x$ onto a semicircle in the plane
Embeddings of the original and transformed data set
Estimation of the drift coefficient in Diffusion Map space
Estimation of the diffusion coefficient in Diffusion Map space
Estimation of the drift coefficient in diffusion map space
Estimation of the diffusion coefficient in diffusion map space
 [1] Min Xi, Wenyu Sun, Jun Chen. Survey of derivative-free optimization. Numerical Algebra, Control & Optimization, 2020, 10 (4) : 537-555. doi: 10.3934/naco.2020050 [2] A. M. Bagirov, Moumita Ghosh, Dean Webb. A derivative-free method for linearly constrained nonsmooth optimization. Journal of Industrial & Management Optimization, 2006, 2 (3) : 319-338. doi: 10.3934/jimo.2006.2.319 [3] Liang Zhang, Wenyu Sun, Raimundo J. B. de Sampaio, Jinyun Yuan. A wedge trust region method with self-correcting geometry for derivative-free optimization. Numerical Algebra, Control & Optimization, 2015, 5 (2) : 169-184. doi: 10.3934/naco.2015.5.169 [4] Jun Takaki, Nobuo Yamashita. A derivative-free trust-region algorithm for unconstrained optimization with controlled error. Numerical Algebra, Control & Optimization, 2011, 1 (1) : 117-145. doi: 10.3934/naco.2011.1.117 [5] Wei-Zhe Gu, Li-Yong Lu. The linear convergence of a derivative-free descent method for nonlinear complementarity problems. Journal of Industrial & Management Optimization, 2017, 13 (2) : 531-548. doi: 10.3934/jimo.2016030 [6] Gaohang Yu. A derivative-free method for solving large-scale nonlinear systems of equations. Journal of Industrial & Management Optimization, 2010, 6 (1) : 149-160. doi: 10.3934/jimo.2010.6.149 [7] Dong-Hui Li, Xiao-Lin Wang. A modified Fletcher-Reeves-Type derivative-free method for symmetric nonlinear equations. Numerical Algebra, Control & Optimization, 2011, 1 (1) : 71-82. doi: 10.3934/naco.2011.1.71 [8] Yigui Ou, Wenjie Xu. A unified derivative-free projection method model for large-scale nonlinear equations with convex constraints. Journal of Industrial & Management Optimization, 2021  doi: 10.3934/jimo.2021125 [9] Aurore Back, Emmanuel Frénod. Geometric two-scale convergence on manifold and applications to the Vlasov equation. Discrete & Continuous Dynamical Systems - S, 2015, 8 (1) : 223-241. doi: 10.3934/dcdss.2015.8.223 [10] Bülent Karasözen. Survey of trust-region derivative free optimization methods. Journal of Industrial & Management Optimization, 2007, 3 (2) : 321-334. doi: 10.3934/jimo.2007.3.321 [11] Yukang He, Zhengwen He, Nengmin Wang. Tabu search and simulated annealing for resource-constrained multi-project scheduling to minimize maximal cash flow gap. Journal of Industrial & Management Optimization, 2021, 17 (5) : 2451-2474. doi: 10.3934/jimo.2020077 [12] Jan Haskovec, Nader Masmoudi, Christian Schmeiser, Mohamed Lazhar Tayeb. The Spherical Harmonics Expansion model coupled to the Poisson equation. Kinetic & Related Models, 2011, 4 (4) : 1063-1079. doi: 10.3934/krm.2011.4.1063 [13] Lan Luo, Zhe Zhang, Yong Yin. Simulated annealing and genetic algorithm based method for a bi-level seru loading problem with worker assignment in seru production systems. Journal of Industrial & Management Optimization, 2021, 17 (2) : 779-803. doi: 10.3934/jimo.2019134 [14] T. W. Leung, Chi Kin Chan, Marvin D. Troutt. A mixed simulated annealing-genetic algorithm approach to the multi-buyer multi-item joint replenishment problem: advantages of meta-heuristics. Journal of Industrial & Management Optimization, 2008, 4 (1) : 53-66. doi: 10.3934/jimo.2008.4.53 [15] Antonios Zagaris, Christophe Vandekerckhove, C. William Gear, Tasso J. Kaper, Ioannis G. Kevrekidis. Stability and stabilization of the constrained runs schemes for equation-free projection to a slow manifold. Discrete & Continuous Dynamical Systems, 2012, 32 (8) : 2759-2803. doi: 10.3934/dcds.2012.32.2759 [16] Amadeu Delshams, Rodrigo G. Schaefer. Arnold diffusion for a complete family of perturbations with two independent harmonics. Discrete & Continuous Dynamical Systems, 2018, 38 (12) : 6047-6072. doi: 10.3934/dcds.2018261 [17] Mauro Maggioni, James M. Murphy. Learning by active nonlinear diffusion. Foundations of Data Science, 2019, 1 (3) : 271-291. doi: 10.3934/fods.2019012 [18] Barbara Kaltenbacher, Gunther Peichl. The shape derivative for an optimization problem in lithotripsy. Evolution Equations & Control Theory, 2016, 5 (3) : 399-430. doi: 10.3934/eect.2016011 [19] Maho Endo, Yuki Kaneko, Yoshio Yamada. Free boundary problem for a reaction-diffusion equation with positive bistable nonlinearity. Discrete & Continuous Dynamical Systems, 2020, 40 (6) : 3375-3394. doi: 10.3934/dcds.2020033 [20] Editorial Office. WITHDRAWN: Fractional diffusion equation described by the Atangana-Baleanu fractional derivative and its approximate solution. Discrete & Continuous Dynamical Systems - S, 2021  doi: 10.3934/dcdss.2020173

Impact Factor:

Tools

Article outline

Figures and Tables