Search

Google Vizier has performed millions of optimizations and accelerated numerous research and production systems at Google, demonstrating the success of Bayesian optimization as a large-scale service. Over multiple years, its algorithm has been improved considerably, through the collective experiences of numerous research efforts and user feedback. In this technical report, we discuss the implementation details and design choices of the current default algorithm provided by Open Source Vizier. Our experiments on standardized benchmarks reveal its robustness and versatility against well-established industry baselines on multiple practical modes., Comment: Google DeepMind Technical Report. Code can be found in https://github.com/google/vizier

This study explores the suitability of semi-insulating compounds, specifically (Cd,Mn)Te and (Cd,Mn)(Te,Se), as materials for room temperature X-ray and gamma-ray detectors. These compounds were grown using the Bridgman method, known for its efficient growth rate. The investigation aims to compare their crystal structure, mechanical properties, optical characteristics, and radiation detection capabilities. The addition of selenium to (Cd,Mn)Te increased the compound's hardness. However, (Cd,Mn)(Te,Se) exhibited one order of magnitude higher etch pit density compared to (Cd,Mn)Te. Photoluminescence analysis at low temperatures revealed the presence of defect states in both materials, characterized by shallow and deep donor-acceptor pair transitions (DAP). Annealing in cadmium vapors effectively eliminated DAP luminescence in (Cd,Mn)Te but not in (Cd,Mn)(Te,Se). Spectroscopic performance assessments indicated that the (Cd,Mn)Te detector outperformed the (Cd,Mn)(Te,Se) detector in responding to a Co-57 source. The reduced performance in the latter case may be attributed to either the presence of a deep trap related to deep DAP luminescence, minimally affected by annealing, or the dominant presence of block-like structures in the samples, as indicated by X-ray diffraction measurements. The block-like structures in (Cd,Mn)(Te,Se) showed ten times larger misorientation angles compared to the (Cd,Mn)Te crystals. (Cd,Mn)Te crystal revealed excellent single crystal properties, demonstrated by narrower omega scan widths. The study also highlights the influence of grain boundaries and twins on crystal structure quality. In our opinion, Bridgman-grown (Cd,Mn)Te shows greater promise as a material for X-ray and gamma-ray detectors compared to (Cd,Mn)(Te,Se)., Comment: 33 pages, 11 figures

Increases in wildfire activity and the resulting impacts have prompted the development of high-resolution wildfire behavior models for forecasting fire spread. Recent progress in using satellites to detect fire locations further provides the opportunity to use measurements to improve fire spread forecasts from numerical models through data assimilation. This work develops a method for inferring the history of a wildfire from satellite measurements, providing the necessary information to initialize coupled atmosphere-wildfire models from a measured wildfire state in a physics-informed approach. The fire arrival time, which is the time the fire reaches a given spatial location, acts as a succinct representation of the history of a wildfire. In this work, a conditional Wasserstein Generative Adversarial Network (cWGAN), trained with WRF-SFIRE simulations, is used to infer the fire arrival time from satellite active fire data. The cWGAN is used to produce samples of likely fire arrival times from the conditional distribution of arrival times given satellite active fire detections. Samples produced by the cWGAN are further used to assess the uncertainty of predictions. The cWGAN is tested on four California wildfires occurring between 2020 and 2022, and predictions for fire extent are compared against high resolution airborne infrared measurements. Further, the predicted ignition times are compared with reported ignition times. An average Sorensen's coefficient of 0.81 for the fire perimeters and an average ignition time error of 32 minutes suggest that the method is highly accurate.

The current fuel moisture content (FMC) subsystems in WRF-SFIRE and its workflow system WRFx use a time-lag differential equation model with assimilation of data from FMC sensors on Remote Automated Weather Stations (RAWS) by the extended augmented Kalman filter. But the quality of the result is constrained by the limitations of the model and of the Kalman filter. We observe that the data flow in a system consisting of a model and the Kalman filter can be interpreted to be the same as the data flow in a recurrent neural network (RNN). Thus, instead of building more sophisticated models and data assimilation methods, we want to train a RNN to approximate the dynamics of the response of the FMC sensor to a time series of environmental data. Because standard AI approaches did not converge to reasonable solutions, we pre-train the RNN with special initial weights devised to turn it into a numerical solver of the differential equation. We then allow the AI training machinery to optimize the RNN weights to fit the data better. We illustrate the method on an example of a time series of 10h-FMC from RAWS and weather data from the Real-Time Mesoscale Analysis (RTMA)., Comment: 4 pages, 4 figures. Seminar on Numerical Analysis SNA'23, Ostrava, Czech Republic, January 23-27, 2023

Vizier is the de-facto blackbox and hyperparameter optimization service across Google, having optimized some of Google's largest products and research efforts. To operate at the scale of tuning thousands of users' critical systems, Google Vizier solved key design challenges in providing multiple different features, while remaining fully fault-tolerant. In this paper, we introduce Open Source (OSS) Vizier, a standalone Python-based interface for blackbox optimization and research, based on the Google-internal Vizier infrastructure and framework. OSS Vizier provides an API capable of defining and solving a wide variety of optimization problems, including multi-metric, early stopping, transfer learning, and conditional search. Furthermore, it is designed to be a distributed system that assures reliability, and allows multiple parallel evaluations of the user's objective function. The flexible RPC-based infrastructure allows users to access OSS Vizier from binaries written in any language. OSS Vizier also provides a back-end ("Pythia") API that gives algorithm authors a way to interface new algorithms with the core OSS Vizier system. OSS Vizier is available at https://github.com/google/vizier., Comment: Published as a conference paper for the systems track at the 1st International Conference on Automated Machine Learning (AutoML-Conf 2022). Code can be found at https://github.com/google/vizier

Meta-learning hyperparameter optimization (HPO) algorithms from prior experiments is a promising approach to improve optimization efficiency over objective functions from a similar distribution. However, existing methods are restricted to learning from experiments sharing the same set of hyperparameters. In this paper, we introduce the OptFormer, the first text-based Transformer HPO framework that provides a universal end-to-end interface for jointly learning policy and function prediction when trained on vast tuning data from the wild, such as Google's Vizier database, one of the world's largest HPO datasets. Our extensive experiments demonstrate that the OptFormer can simultaneously imitate at least 7 different HPO algorithms, which can be further improved via its function uncertainty estimates. Compared to a Gaussian Process, the OptFormer also learns a robust prior distribution for hyperparameter response functions, and can thereby provide more accurate and better calibrated predictions. This work paves the path to future extensions for training a Transformer-based model as a general HPO optimizer., Comment: Published as a conference paper in Neural Information Processing Systems (NeurIPS) 2022. Code can be found in https://github.com/google-research/optformer and Google AI Blog can be found in https://ai.googleblog.com/2022/08/optformer-towards-universal.html

We present a simple finite element formulation of mass-consistent approximation, and a fast multigrid iterative method with adaptive semicoarsening, which maintains the convergence of the iterations over a range of grids and penalty coefficients. The method is designed to run in each time step of WRF-SFIRE and replace the interpolation from the atmospheric mesh to the fire mesh by a diagnostic model of the air flow over a fine-scale terrain., Comment: 4 pages, Seminar on Numerical Analysis - SNA'21, Ostrava, Czech Republic, January 25--28, 2021

Predicting changes in sea ice cover is critical for shipping, ecosystem monitoring, and climate modeling. Current sea ice models, however, predict more ice than is observed in the Arctic, and less in the Antarctic. Improving the fit of these physics-based models to observations is challenging because the models are expensive to run, and therefore expensive to optimize. Here, we construct a machine learning surrogate that emulates the effect of changing model physics on forecasts of sea ice area from the Los Alamos Sea Ice Model (CICE). We use the surrogate model to investigate the sensitivity of CICE to changes in the parameters governing: ice's ridging and albedo; snow's albedo, aging, and thermal conductivity; the effect of meltwater on albedo; and the effect of ponds on albedo. We find that CICE's sensitivity to these model parameters differs between hemispheres. We propose that future sea ice modelers separate the snow conductivity and snow grain size distributions on a seasonal and inter-hemispheric basis, and we recommend optimal values of these parameters. This will make it possible to make models that fit observations of both Arctic and Antarctic sea ice more closely. These results demonstrate that important aspects of the behavior of a leading sea ice model can be captured by a relatively simple support vector regression surrogate model, and that this surrogate dramatically increases the ease of tuning the full simulation., Comment: 6 pages

A dry decade in the Navajo Nation has killed vegetation, dessicated soils, and released once-stable sand into the wind. This sand now covers one-third of the Nation's land, threatening roads, gardens and hundreds of homes. Many arid regions have similar problems: global warming has increased dune movement across farmland in Namibia and Angola, and the southwestern US. Current dune models, unfortunately, do not scale well enough to provide useful forecasts for the $\sim$5\% of land surfaces covered by mobile sand. We test the ability of two deep learning algorithms, a GAN and a CNN, to model the motion of sand dunes. The models are trained on simulated data from community-standard cellular automaton model of sand dunes. Preliminary results show the GAN producing reasonable forward predictions of dune migration at ten million times the speed of the existing model., Comment: Workshop on Tackling climate change with machine learning at NeurIPS. Vancouver, Canada, December 2019

Zeroth-order optimization is the process of minimizing an objective $f(x)$, given oracle access to evaluations at adaptively chosen inputs $x$. In this paper, we present two simple yet powerful GradientLess Descent (GLD) algorithms that do not rely on an underlying gradient estimate and are numerically stable. We analyze our algorithm from a novel geometric perspective and present a novel analysis that shows convergence within an $\epsilon$-ball of the optimum in $O(kQ\log(n)\log(R/\epsilon))$ evaluations, for any monotone transform of a smooth and strongly convex objective with latent dimension $k < n$, where the input dimension is $n$, $R$ is the diameter of the input space and $Q$ is the condition number. Our rates are the first of its kind to be both 1) poly-logarithmically dependent on dimensionality and 2) invariant under monotone transformations. We further leverage our geometric perspective to show that our analysis is optimal. Both monotone invariance and its ability to utilize a low latent dimensionality are key to the empirical success of our algorithms, as demonstrated on BBOB and MuJoCo benchmarks., Comment: 11 main pages, 26 total pages

Climate change is one of the greatest challenges facing humanity, and we, as machine learning experts, may wonder how we can help. Here we describe how machine learning can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by machine learning, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the machine learning community to join the global effort against climate change., Comment: For additional resources, please visit the website that accompanies this paper: https://www.climatechange.ai/

We analyze the accuracy of traffic simulations metamodels based on neural networks and gradient boosting models (LightGBM), applied to traffic optimization as fitness functions of genetic algorithms. Our metamodels approximate outcomes of traffic simulations (the total time of waiting on a red signal) taking as an input different traffic signal settings, in order to efficiently find (sub)optimal settings. Their accuracy was proven to be very good on randomly selected test sets, but it turned out that the accuracy may drop in case of settings expected (according to genetic algorithms) to be close to local optima, which makes the traffic optimization process more difficult. In this work, we investigate 16 different metamodels and 20 settings of genetic algorithms, in order to understand what are the reasons of this phenomenon, what is its scale, how it can be mitigated and what can be potentially done to design better real-time traffic optimization methods., Comment: Presented at NeurIPS 2018 Workshop "Machine Learning for Intelligent Transportation Systems" https://sites.google.com/site/nips2018mlits

Data likelihood of fire detection is the probability of the observed detection outcome given the state of the fire spread model. We derive fire detection likelihood of satellite data as a function of the fire arrival time on the model grid. The data likelihood is constructed by a combination of the burn model, the logistic regression of the active fires detections, and the Gaussian distribution of the geolocation error. The use of the data likelihood is then demonstrated by an estimation of the ignition point of a wildland fire by the maximization of the likelihood of MODIS and VIIRS data over multiple possible ignition points., Comment: 12 pages, 6 figures; VIII International Conference on Forest Fire Research, Coimbra, Portugal, November 2018

Observational data collected during experiments, such as the planned Fire and Smoke Model Evaluation Experiment (FASMEE), are critical for progressing and transitioning coupled fire-atmosphere models like WRF-SFIRE and WRF-SFIRE-CHEM into operational use. Historical meteorological data, representing typical weather conditions for the anticipated burn locations and times, have been processed to initialize and run a set of simulations representing the planned experimental burns. Based on an analysis of these numerical simulations, this paper provides recommendations on the experimental setup that include the ignition procedures, size and duration of the burns, and optimal sensor placement. New techniques are developed to initialize coupled fire-atmosphere simulations with weather conditions typical of the planned burn locations and time of the year. Analysis of variation and sensitivity analysis of simulation design to model parameters by repeated Latin Hypercube Sampling are used to assess the locations of the sensors. The simulations provide the locations of the measurements that maximize the expected variation of the sensor outputs with the model parameters., Comment: 35 pages, 4 tables, 28 figures

Assimilation of data into a fire-spread model is formulated as an optimization problem. The level set equation, which relates the fire arrival time and the rate of spread, is allowed to be satisfied only approximately, and we minimize a norm of the residual. Previous methods based on modification of the fire arrival time either used an additive correction to the fire arrival time, or made a position correction. Unlike additive fire arrival time corrections, the new method respects the dependence of the fire rate of spread on diurnal changes of fuel moisture and on weather changes, and, unlike position corrections, it respects the dependence of the fire spread on fuels and terrain as well. The method is used to interpolate the fire arrival time between two perimeters by imposing the fire arrival time at the perimeters as constraints., Comment: 13 pages, 6 figures, ICCS 2018. Minor update

A new Thurstonian rating scale model uses a variable decision rule (VDR) that incorporates three previously formulated, distinct decision rules. The model includes probabilities for choosing each rule, along with Gaussian representation and criterion densities. Numerical optimisation techniques were validated through demonstrating that the model fits simulated data tightly. For simulations with 400 trials per stimulus (tps), useful information emerged about the generating parameters. However, larger experiments (e.g. 4000 tps) proved desirable for better recovery of generating parameters and to support trustworthy choices between competing models by the Akaike Information Criterion. In reanalyses of experiments by others, the VDR model explained most of the data better than did classical signal detection theory models., Comment: Contains source code as an attachment to this PDF, in files 2015-12-02_speechresearch.tgz and Rosner_Kochanski_2016.tgz