Interactive Sonification Monitoring in Evolutionary Optimization

This case study introduces interactive sonification to evolutionary strategies (ES) for global optimization. We briefly describe the specific strengths of sonification as a tool for monitoring, the emerging trend of interactive sonification, and what it can add to the field of evolutionary computation. Then we line out the background of ES as optimization heuristics, briefly explain the algorithmic procedure of ES and discuss the need to intervene during optimization runs and the current shortcomings in appropriate user feedback. This motivates the development of an auditory closed loop setup that brings the expertise of interactive sonification to the field of monitoring ES algorithms. Further, we present considerations for the sound design and the detailed mapping of parameters from the ES to sound properties. Finally, we discuss the various implemented modes of interaction and their significance for the optimization through ES. 1. MONITORING THROUGH SONIFICATION In all the different fields of applications, monitoring is amongst the most suitable ones for the use of sonification. Well-established examples range from the operating theatre to monitoring seismograms by listening [1]. Sonification has further been used for the monitoring of stock markets [2], network traffic [3], electrocardiograms [4], quantum oscillations [5], and EEG data [6][7]. The widespread use as a monitoring tool is because the human auditory system is particularly apt for this task. The two most important listening abilities for the purpose of monitoring are backgrounding, which sets in when a sound becomes steady, as well as the ability to focus on selected streams in a mixture of sounds [8]. Additionally, the auditory system has a strong capacity to readily notice transient sounds. Finally, and most important for monitoring, the user does not need to have a particular orientation in space in order to follow a process by listening. 2. FIELD OF APPLICATION In our work we apply interactive sonification to evolution strategies (ES), which have grown into powerful optimization heuristics [9]. ES algorithms are biologically inspired, population based, randomized search heuristics. ES apply the principles of biological evolution to optimization: inheritance and mutation of genes, and selection of the fittest solutions according to the famous Darwinian principle. In the sixties and seventies, Fogel [10], Holland [11], Rechenberg [12] and Schwefel [13] translated these paradigms into algorithms that are called evolutionary computation today, a field that has evolved and diversified into a rich and frequently used set of methods for optimization problems. The resulting algorithms search efficiently for optimal solutions in high dimensional parameter spaces. 2.1. The principles of ES For a better understanding of the sound design, which will be described later, we briefly introduce in more detail the underlying principle of the optimization procedure. The aim of finding an optimal solution corresponds to finding the global minimum of a cost-function, which is usually embedded in a high dimensional parameter space. ES are particularly useful in black black box optimization scenarios, e.g., when no derivates are available, which could be invested into the search process. This is why ES contain a random element in exploring the search space. • The initial step is to randomly select a set (population) of points (individuals) that covers the search space of interest. • Secondly, their fitness, which corresponds to the value of the cost-function, is evaluated. • In a third step a defined percentage of the species with the poorest fitness values are discarded (selection). • Forth, offspring is produced by new species that are derived from the fittest by varying their position in parameter space with a certain mutation strength (distribution σ) around their ancestors (inheritance). The 17th International Conference on Auditory Display (ICAD-2011) June 20-24, 2011, Budapest, Hungary This procedure is repeated until the selection of points converges towards the optimal solution. The development of the cost function and the σ vector for examples of a converging and a non-converging optimization is shown in figure 1.