This page provides access to our implementation of StoSOO (Stochastic Simultaneous Optimistic Optimization). StoSOO was introduced in this Valko, Carpentier, Munos, ICML 2013 paper.
We named this version of StoSOO yaStoSOO which stands for "yet another StoSOO".
StoSOO is a black-box global optimization algorithm. StoSOO makes almost no assumption about the function being optimized (in particular, StoSOO does not assume differentiability).
StoSOO optimizes a real function which is stochastic (each time a point is evaluated, a different value is returned). It may also be used to optimize deterministic functions under the guise of SOO, an algorithm studied in this NIPS 2011 paper by Rémi Munos.
StoSOO and SOO hence fall into the "global optimizer" category.
Originally, these algorithms were designed as an application of bandit theory to function optimization. Hence, the algorithm may be seen as very pure; as function optimizers, they obviously lack many features that should be introduced to improve their performance. Anyway, StoSOO and SOO have very appealing theoretical properties regarding their expected performance after a certain amount of function evaluations; as far as we know, they are the only global otpimization algorithms for which such properties have been prooved.
StoSOO and SOO work by partitioning, in the same spirit as Jones's DiRect.
Download: The most recent release is v.0.3. It is available by clicking on this link. I'd appreciate that you send me an email when you download the software (philippe -dot- preux -at- univ-lille3 -dot- fr); that way, I'll be able to inform you of any future release. I'd also be very happy to hear about your use of this software.
This implementation is in C, and has been developed under Linux/Ubuntu. Its installation on any Linux system, and Mac OS, is probably very easy and straitforward. I do not see any reason why this code would'nt run on other platform, but I do not intend to care about it.
Manual: There is a manual about the installation and how to use this code. It is available by clicking on this link.
Experimental assessment of SOO:
History of versions:
Please note that this software has been extensively debugged while optimizing deterministic functions (hence, in SOO mode). We will currently paying strong attention to noisy optimization to assess StoSOO.