Neural architecture search (NAS) methods rely on a search strategy for deciding which architectures to evaluate next and a performance estimation strategy for assessing their performance (e.g., using full evaluations, multi-fidelity evaluations, or the one-shot model). In this paper, we focus on the search strategy. We introduce the simple yet powerful evolutionary algorithm of differential evolution to the NAS community. Using the simplest performance evaluation strategy of full evaluations, we comprehensively compare this search strategy to regularized evolution and Bayesian optimization and demonstrate that it yields improved and more robust results for 13 tabular NAS benchmarks based on NAS-Bench-101, NAS-Bench-1Shot1, NAS-Bench-201 and NAS-HPO bench.
First-order stochastic methods for solving large-scale non-convex optimization problems are widely used in many big-data applications, e.g. training deep neural networks as well as other complex and potentially non-convex machine learning models. Their inexpensive iterations generally come together with slow global convergence rate (mostly sublinear), leading to the necessity of carrying out a very high number of iterations before the iterates reach a neighborhood of a minimizer. In this work, we present a first-order stochastic algorithm based on a combination of homotopy methods and SGD, called Homotopy-Stochastic Gradient Descent (H-SGD), which finds interesting connections with some proposed heuristics in the literature, e.g. optimization by Gaussian continuation, training by diffusion, mollifying networks. Under some mild assumptions on the problem structure, we conduct a theoretical analysis of the proposed algorithm. Our analysis shows that, with a specifically designed scheme for the homotopy parameter, H-SGD enjoys a global linear rate of convergence to a neighborhood of a minimum while maintaining fast and inexpensive iterations. Experimental evaluations confirm the theoretical results and show that H-SGD can outperform standard SGD.
After developer adjustments to a machine learning (ML) algorithm, how can the results of an old hyperparameter optimization (HPO) automatically be used to speedup a new HPO? This question poses a challenging problem, as developer adjustments can change which hyperparameter settings perform well, or even the hyperparameter search space itself. While many approaches exist that leverage knowledge obtained on previous tasks, so far, knowledge from previous development steps remains entirely untapped. In this work, we remedy this situation and propose a new research framework: hyperparameter transfer across adjustments (HT-AA). To lay a solid foundation for this research framework, we provide four simple HT-AA baseline algorithms and eight benchmarks changing various aspects of ML algorithms, their hyperparameter search spaces, and the neural architectures used. The best baseline, on average and depending on the budgets for the old and new HPO, reaches a given performance 1.2--2.6x faster than a prominent HPO algorithm without transfer. As HPO is a crucial step in ML development but requires extensive computational resources, this speedup would lead to faster development cycles, lower costs, and reduced environmental impacts. To make these benefits available to ML developers off-the-shelf and to facilitate future research on HT-AA, we provide python packages for our baselines and benchmarks.
The development of domain-independent planners within the AI Planning community is leading to "off-the-shelf" technology that can be used in a wide range of applications. Moreover, it allows a modular approach --in which planners and domain knowledge are modules of larger software applications-- that facilitates substitutions or improvements of individual modules without changing the rest of the system. This approach also supports the use of reformulation and configuration techniques, which transform how a model is represented in order to improve the efficiency of plan generation. In this article, we investigate how the performance of domain-independent planners is affected by domain model configuration, i.e., the order in which elements are ordered in the model, particularly in the light of planner comparisons. We then introduce techniques for the online and offline configuration of domain models, and we analyse the impact of domain model configuration on other reformulation approaches, such as macros.
In this paper, we propose an approach to neural architecture search (NAS) based on graph embeddings. NAS has been addressed previously using discrete, sampling based methods, which are computationally expensive as well as differentiable approaches, which come at lower costs but enforce stronger constraints on the search space. The proposed approach leverages advantages from both sides by building a smooth variational neural architecture embedding space in which we evaluate a structural subset of architectures at training time using the predicted performance while it allows to extrapolate from this subspace at inference time. We evaluate the proposed approach in the context of two common search spaces, the graph structure defined by the ENAS approach and the NAS-Bench-101 search space, and improve over the state of the art in both.
Despite significant progress in challenging problems across various domains, applying state-of-the-art deep reinforcement learning (RL) algorithms remains challenging due to their sensitivity to the choice of hyperparameters. This sensitivity can partly be attributed to the non-stationarity of the RL problem, potentially requiring different hyperparameter settings at various stages of the learning process. Additionally, in the RL setting, hyperparameter optimization (HPO) requires a large number of environment interactions, hindering the transfer of the successes in RL to real-world applications. In this work, we tackle the issues of sample-efficient and dynamic HPO in RL. We propose a population-based automated RL (AutoRL) framework to meta-optimize arbitrary off-policy RL algorithms. In this framework, we optimize the hyperparameters and also the neural architecture while simultaneously training the agent. By sharing the collected experience across the population, we substantially increase the sample efficiency of the meta-optimization. We demonstrate the capabilities of our sample-efficient AutoRL approach in a case study with the popular TD3 algorithm in the MuJoCo benchmark suite, where we reduce the number of environment interactions needed for meta-optimization by up to an order of magnitude compared to population-based training.
In many fields of study, we only observe lower bounds on the true response value of some experiments. When fitting a regression model to predict the distribution of the outcomes, we cannot simply drop these right-censored observations, but need to properly model them. In this work, we focus on the concept of censored data in the light of model-based optimization where prematurely terminating evaluations (and thus generating right-censored data) is a key factor for efficiency, e.g., when searching for an algorithm configuration that minimizes runtime of the algorithm at hand. Neural networks (NNs) have been demonstrated to work well at the core of model-based optimization procedures and here we extend them to handle these censored observations. We propose (i)~a loss function based on the Tobit model to incorporate censored samples into training and (ii) use an ensemble of networks to model the posterior distribution. To nevertheless be efficient in terms of optimization-overhead, we propose to use Thompson sampling s.t. we only need to train a single NN in each iteration. Our experiments show that our trained regression models achieve a better predictive quality than several baselines and that our approach achieves new state-of-the-art performance for model-based optimization on two optimization problems: minimizing the solution time of a SAT solver and the time-to-accuracy of neural networks.
Neural Architecture Search (NAS) is a logical next step in the automatic learning of representations, but the development of NAS methods is slowed by high computational demands. As a remedy, several tabular NAS benchmarks were proposed to simulate runs of NAS methods in seconds. However, all existing NAS benchmarks are limited to extremely small architectural spaces since they rely on exhaustive evaluations of the space. This leads to unrealistic results, such as a strong performance of local search and random search, that do not transfer to larger search spaces. To overcome this fundamental limitation, we propose NAS-Bench-301, the first model-based surrogate NAS benchmark, using a search space containing $10^{18}$ architectures, orders of magnitude larger than any previous NAS benchmark. We first motivate the benefits of using such a surrogate benchmark compared to a tabular one by smoothing out the noise stemming from the stochasticity of single SGD runs in a tabular benchmark. Then, we analyze our new dataset consisting of architecture evaluations and comprehensively evaluate various regression models as surrogates to demonstrate their capability to model the architecture space, also using deep ensembles to model uncertainty. Finally, we benchmark a wide range of NAS algorithms using NAS-Bench-301 allowing us to obtain comparable results to the true benchmark at a fraction of the cost.
Automated Machine Learning, which supports practitioners and researchers with the tedious task of manually designing machine learning pipelines, has recently achieved substantial success. In this paper we introduce new Automated Machine Learning (AutoML) techniques motivated by our winning submission to the second ChaLearn AutoML challenge, PoSH Auto-sklearn. For this, we extend Auto-sklearn with a new, simpler meta-learning technique, improve its way of handling iterative algorithms and enhance it with a successful bandit strategy for budget allocation. Furthermore, we go one step further and study the design space of AutoML itself and propose a solution towards truly hand-free AutoML. Together, these changes give rise to the next generation of our AutoML system, Auto-sklearn (2.0). We verify the improvement by these additions in a large experimental study on 39 AutoML benchmark datasets and conclude the paper by comparing to Auto-sklearn (1.0), reducing the regret by up to a factor of five.