Community Research and Development Information Service - CORDIS

Optimising optimisation algorithms

In many modern problems and, in particular, in those arising in machine learning, the amount of data is too large to apply standard optimisation algorithms. EU-funded scientists developed new algorithms that rely on a fraction of the input data to reduce the running time.
Optimising optimisation algorithms
Internet traffic logs and financial transactions are just two examples of massive data sets that appear more and more often in various applications. Analysing and managing such data sets force scientists to reconsider the conventional approaches to developing efficient optimisation algorithms.

Optimisation algorithms are used to evaluate design trade-offs and find patterns in data sets. Machine-learning algorithms often attempt to identify features that help in classification tasks. Finding the smallest set of features with the greatest predictive value is an optimisation problem.

Scientists working on the EU-funded project SUBLINEAROPTML (Sublinear optimization for machine learning) developed algorithms that run in sublinear time for solving such optimisation problems. These are based on a combination of advanced sampling techniques and a randomised implementation of online learning algorithms.

Online learning algorithms make a prediction for each element in a stream of data, and with the feedback received their accuracy is improved for subsequent predictions. In contrast to statistical machine learning, they don't make assumptions about the input data.

The new algorithms make use of randomisation to prune data and produce correct solutions despite running in time that is smaller than the data representation – the so-called sublinear time. Project scientists have shown that the running times of most of these algorithms are the best possible in random access machine (RAM) models.

Project algorithms can also be extended to kernelised versions of these problems, including support vector machine optimisation problems for which sublinear-time solvers were not available. These advances in machine learning have been presented at several international conferences.

SUBLINEAROPTML has resulted in a potentially more efficient way to help computers solve some of the toughest optimisation problems they face. A router may one day utilise the new algorithms to calculate the fastest path through a busy network.

Related information

Keywords

Optimisation algorithms, machine learning, massive data sets, optimisation problem, learning algorithms
Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top