Login
The UniverSelf project is proud to announce the release of its first in a series of technical reports on:
Deliverable D3.1 - Identification of suitable classes of methods for parameter optimization.
The report can be accessed under the Dissemination/Technical Reports section, or directly at the following link http://www.univerself-project.eu/technical-reports
Summary:
This deliverable focuses on network optimization tasks and the selection of methods suitable to achieve these tasks. The corresponding question we would like to answer is how to choose changes in network parameters that give us an improvement for a given set of key performance indicators. In principle there are two types of problems: those where a model is available giving us an estimate what a parameter change would approximately result in and those where there is no such model and where the best parameter changes can only be guessed.
The first chapter focuses on two classes of methods where no model is available, i.e. changes in parameters will result in changes of the key performance indicators that are hardly or not at all predictable. In this context the role of randomness in two completely disjunctive classes of algorithms, namely evolutionary algorithms representing heuristic methods on the one hand and a gradient descent approach representing a solution designed for convex optimization problems on the other hand, is analysed. We show that evolutionary algorithms can be successfully applied for non-convex problems where the search space is large; a major advantage of this class of algorithms is its flexibility – a telecom researcher can choose from a wide variety of different flavours of the method class (e.g. genetic algorithms, genetic programming). In this deliverable we analyse evolutionary algorithms as a placeholder for methods with random elements. For the specific example of multihop relay-assisted cellular networks, we have clearly shown that evolutionary algorithms outperform methods like simulated annealing as well as mixed integer linear programming. The major advantage of evolutionary algorithms is the fact that a solution can be found relatively quickly and that the state space is explored comprehensively so that ending up in a local optimum becomes less likely. The next steps of our work in this field will be the investigation of the suitability for optimizations within already running networks where trying out unfavourable parameter configurations may have a severe impact.
By contrast, gradient descent approaches are typically applied to convex problems. However, wireless access networks, due to their statistical and non-deterministic properties, reveal problems that are only nearly, but not strictly convex. In this case, we propose to adapt the gradient descent approach by injecting a certain level of noise, similar to the evolutionary algorithms, such that the solution does not get stuck in local optima.
The following chapters then focus more on use cases rather than on the methods itself. The chapter on Governance (use case 6 in deliverable D4.1) extends the method-oriented view by an integrated view on the network. Instead of solving problems for a wireless access network alone, it reveals an additional dimension of complexity by looking at both the wireless and the core domains of the network in an end-to-end way. We show that this complexity can successfully be addressed by using policies that control and coordinate the performance of the entire network, not only individual network domains. Also, there are concrete instances where evolutionary algorithms are applied. This chapter addresses a unification/federation aspect and thus acts as a link to the work on UMF (Unified Management Framework) (deliverable D2.1). With respect to state-of-the-art work, this chapter focuses on a joint end-to-end management of networks composed by multi-vendor/multi-technology segments whose governance is policy-based. Additional effort will have to be spent to answer the question what specific methods are best-suited for this particular purpose.
The next chapter covers a large variety of different facets of load balancing (mainly referring to use case 3 in deliverable D4.1). Load balancing techniques can namely be used in the access domain of the network, in the backhaul/core domain or at the interface of both. A load balancing framework was developed integrating all these aspects in a global view. The analysed aspects also include load balancing between different radio access technologies, interference coordination, and transmission power adaptation as well as access point (de-) activation. The presented work exceeds state-of-the-art work by, e.g., taking into account aspects like energy efficiency or by combining existing load balancing approaches for a better performance. This work bears significant potential, as load balancing is viewed from a multitude of perspectives. The final integration, also in terms of best-suited methods, of these related pieces into one mosaic will be the work of the second project year.
In conclusion, this deliverable elaborates on a class of methods that is particularly eligible for non-deterministic situations (as they appear in wireless access networks). Furthermore, solutions for two particular fields are proposed which both require an integrated view on the network: first of all governance spanning over several network domains (wireless and wireline) and secondly load balancing spanning over several network domains (access, backhaul, and core).
