Objetive 1. Analysis, design, implementation and optimization of heterogeneous algorithms

This objective considers all the development of algorithms that can be used as enabling technology. It covers all types of decompositions (NNMF, LU , QR, SVD, etc.), optimization problems, etc.

The algorithms will be implemented, optimized and presented as open-source packages with the objective that the scientific community uses, expands and validates. The developments must be capable of being executed in any type of architecture: CPUs, graphics accelerators (GPUs) or co-processors such as Intel Xeon PHI. And they must use all available resources (heterogeneity) simultaneously.

Therefore, the focus is on maximizing the performance of large computer systems required for HPC problems.

Objetive 2. Development of competitive computational solutions

This objetive is shared with SSPressing-Vound and it is mainly oriented to mobile devices (smartphones, tablets, etc.) and to computational aspects of potential apps to be built. The objective 2 links directly to objective 1 and they are not independent. In the objective 1, the performance of large systems is prioritized and in the objective 2 the energy efficiency is prioritized and, if it is possible, the performance also.

The objective 2 encompasses the entire software development cycle of systems based on low-consumption devices and high mobility that can cooperate in solving the problems. Today there are a big variety of devices with a low consumption and an affordable price, such as Raspberri Pi, Arduino, Nvidia Jetson, etc. They are also smartphones, wearables or tablets. All of them have interesting computing capabilities and often share the architecture of their processor (usually ARM multicore processors), but not the OS or the middleware.

This objetive will focus on providing solutions for these types of hardware. In summary, the objectives 1 and 2 will provide computational solutions (hardware/software) necessary for the other subprojects.

Objetive 3. Study, analysis and application of multi-criteria evolutionary learning algorithms or based on preferences

In many problems there are different quality criteria that must be optimized simultaneously. In many cases the criteria are conflicting and we should seek a compromise. If all criteria can be quantified by numbers, it is possible to define a partial order relation between the different solutions. In the absence of a total order, the learning algorithm will provide a set of minimal elements of that partial order, known as "set of non-dominated solutions". The final solution to the problem is obtained by applying a second algorithm of multi-criteria decision (MCDM, Multi-Criteria Decision Making) to the set of non-dominated solutions.

Recent works in machine learning is working on some extensions of the above problem, based on certain relationships of partial preference between solutions. Multi-criteria optimization problem extends to cases where not all the criteria may be expressed by numbers. One of the most successful learning strategies through preferences is based on the use of metaheuristics and within them, the evolutionary algorithms with inaccurate or symbolic fitness. Therefore, studies will be conducted to find the best fuzzy measures. Emphasis will be placed on the use of new combination operators and random generation techniques of fuzzy measures.