Handbook on optimization, learning, operation and cooperation methods
- Ewe L.
- Gruber M.
This deliverable concludes the work of the work package Network Empowerment with a quantitative comparison of a selection of methods, complemented by a qualitative, handbook-like conclusion on a reasonable classification of methods and recommendations on their suitability for telecom-relevant problems. The work package had originally been designed to capture three different super-classes of methods, namely those concerned with the optimization of network configuration parameters with a certain search space (Parameter Optimization), those concerned with the gathering of information and knowledge to take appropriate decisions that are often discrete (Observation and Action), and those that are on a meta-level covering multiple network segments, domains, or simply several nodes of the same type (Cooperation Strategies and Incentives). The latter display a larger complexity as corresponding algorithms are not confined to a specific location so that communication aspects between distributed instances of the same algorithm need to be considered as well. It is apparent that bio-inspired algorithms have gained a considerable degree of popularity during the last years, especially for problems covered in Parameter Optimization. Bio-inspired algorithms have the advantage that they intrinsically work as proven by nature, whether it is genetic algorithms mimicking the evolution of life or e.g. particle swarm optimization mimicking the emerging cooperation in a larger group of entities. One big challenge an engineer will have when working with this class of algorithms is their inherent complexity. As a matter of fact, a genetic algorithm itself has a number of parameters that need to be tuned (population size, recombination rate, mutation rate, reproduction rate...) and the proper tuning of the parameters may, to a large degree depend on the nature of the problem. This substantially increases the problem complexity a telecom engineer is confronted with. Besides the bio-inspired algorithms, there is a class of deterministic algorithms that is easier to use in this sense. Gradient-ascent algorithms, for instance, are in principle straightforward to understand and use as they try to find the most promising neighbouring solution. Gradient ascent algorithms are problematic in a different sense though. As they focus their attention to immediate neighbours in the search space, it is quite possible that the solution will converge in a local optimum. In this respect, the main difference to the bio-inspired algorithms is the lack of a random element mechanism that allows jumping out of locally optimal solutions. For Observation and Action, one can roughly classify learning algorithms in three classes: supervised learning algorithms where the training data is actually labelled so that the data can be differentiated according to some desired pattern, unsupervised learning algorithms where there is no concrete pattern one is searching for and where the structure is hidden (clustering of data is a typical example for this), and reinforcement learning that can also be seen as a Parameter Optimization method (the solution is iteratively approached making use of a reward system). Besides these learning algorithms, the deliverable also considers more classical and Bayes-derived statistical procedures in order to improve a decision process. For Cooperation Strategies and Incentives, the problem space is much more complex than for the latter two tasks. The reason is that an algorithm is not only responsible for the optimization of one node or entity, but rather for the global optimization of a multitude of domains, segments, nodes or entities. It appears that there is no obvious and straightforward solution to this problem that could be applied and tailored to a variety of problems in a simple way. Unfortunately, each solution needs to be tailored to the individual particularities of the given problem. More global solutions to Cooperation Strategies and Incentives are also not necessarily algorithm-like, but rather architecture- or concept-like. Taking the example of wireless access networks where many SON (self-organizing networks) features can jeopardize the stability of the entire network. The most viable solution seems to be a bipartite concept where separating conflicting features in time wherever possible and, complementarily, jointly optimizing them wherever this is not possible, seems to be the best solution. Generally, when choosing the most suitable method for a given telecom problem, an engineer should consider the following features of an algorithm that all play a substantial role: complexity of the problem, required stability, ability to be run online, algorithmic flexibility regarding parameter search space, implementation effort / simplicity of algorithm, number of parameters that need to be tuned within the algorithm, and convergence time. This work is complemented by a Mapping Modelling Methodolo gy that ext ends the conventional design of functionality by separating design concerns into decision process, main algorithm[s], and driving model.