5 Addressing Optimization Issues in Network Planning with Evolutionary Computation John Tindle and K. F. Poon 5.1 Introduction In line with the rapid growth of telecommunications networks in recent years, there has been a corresponding increase in the level of network complexity. Consequently, it is now generally accepted that advanced computer aided simulation and analysis methods are essential aids to the management of large networks. The 1990s will be recalled as the decade of business process re-engineering. Most large conventional manufacturing organisations have already applied some re-engineering to improve their overall process efficiency, resulting in the establishment of streamlined management structures, optimised production schedules and facilities, reduced staffing levels and much closer control over working budgets and expenditure for major projects. A similar process of re-engineering is now taking place in the telecommunications sector. In the future, information will assume a much greater level of importance. It is becoming increasingly evident that many businesses will achieve and actively maintain their competitive edge by continuously gathering, analysing and exploiting operational data in new and novel ways. Large-scale systems models can now be created and intelligently evolved to produce near-optimal topological structures. Macro-level network system analysis can now be attempted because of developments in heuristic, mathematical and intelligent methods, and the availability of low cost high performance desktop computers. Personal workstations can now solve complex problems by processing vast amounts of data Telecommunications Optimization: Heuristic and Adaptive Techniques, edited by D. Corne, M.J. Oates and G.D. Smith © 2000 John Wiley & Sons, Ltd Telecommunications Optimization: Heuristic and Adaptive Techniques. Edited by David W. Corne, Martin J. Oates, George D. Smith Copyright © 2000 John Wiley & Sons Ltd ISBNs: 0-471-98855-3 (Hardback); 0-470-84163X (Electronic) Telecommunications Optimization: Heuristic and Adaptive Techniques80 using new smart algorithms, thereby providing better solutions and deeper insight into the complex structure of network problems. The primary goals for telecommunications network operators are to develop methodologies for the following: Optimal planning methods are required to minimise the cost of building new network structures. Impact analysis is required to predict the impact that new components and systems will have on the network and business. Business modelling is required to evaluate the effects that selected, individual and grouped changes will ultimately have upon the whole business and organisation. Fault analysis is required to accurately identify and repair faults with the minimum disruption to network performance. This chapter considers network planning and impact analysis for the Plain Old Telephone System (POTS). In the past, new network designs were predominately undertaken by human planners using simple manual or partially automated methods. The methods employed were characteristically labour-intensive with relatively low levels of productivity. A planner, for example, would usually be expected to design a typical network layout within three working days. Generally, this work would entail two stages or more of design iteration, each stage taking twelve hours. In most cases, the solutions generated were sub-optimal, and therefore the resultant capital costs were often higher than necessary. It became clear that there was an urgent need for new automated planning tools to handle complexity and aid network designers. Henderson (1997) provides a very interesting and illustrative case study. This project involved the installation of a national SDH network for Energis UK. He states that, “Building of the network commenced before the design was complete. The truth is, the design was not complete until some time after it was built and working! In practice parallel design and build is becoming more common as the time scale pressures of our competitive environment increase”. In addition, “Unfortunately, build never occurs as planned. Unforeseen difficulties delay sections, whilst others prove easy. With a virtual workforce of 1500 people, working to very tight time scales, logistics control was a major problem”. The inability to produce good plans sufficiently rapidly can sometimes result in poorly assigned capacity levels in a network. It is also likely that sub-optimal designs could be implemented which incur higher than necessary capital costs. In most cases, the additional costs of rework to solve these problems will be very high. 5.2 The Access Network In the UK access, 90% of new systems are copper based and only 10% optical fibre. Incumbent service operators are finding it difficult to justify the installation of fibre as a replacement technology to support narrow band services carried by the existing copper network. At present, the cost of optical network components is considered too high, especially the electronic Line Termination Unit (LTU). In addition, the introduction of Asymmetric Digital Subscriber Line (ADSL) technology has the potential to greatly Addressing Optimization Issues in Network Planning with Evolutionary Computation 81 increase the available bandwidth in the copper access network, thus potentially extending its working lifetime. Consequently, copper bearer platforms still attract significant levels of investment from many telecommunications operators. Although it was previously considered that fibre would be deployed more aggressively in the access sector, copper remains the dominant technology option for the narrow band service sector. The continuous use of copper into the near future has justified the development of a new smart planning tool to automate the planning process for greenfield developments. Although in this instance, the methods described in this chapter have been applied to solve the Copper Planning Problem (CPP), it is believed the approach in general can be successfully tailored to suit a variety of different network problems. In the UK, an ongoing development programme is in place to provide telephony to new housing estates. It is essential that an estimate of the cost of work is produced, a bill of materials and a resource schedule. This information is passed on to construction team for installation. To produce a satisfactory project plan for the construction team by the manual method proved very difficult, time consuming and tedious because of the many complex factors planners were required to consider. The smart tool, GenOSys, which has been developed primarily to aid network planners, automates the design of the secondary cable and duct networks necessary to connect new customers to the exchange. GenOSys employs Evolutionary Computation (EC) techniques to facilitate the rapid solution of large and complex network problems. In order to develop an efficient search strategy yielding optimal or near-optimal solutions, it is essential to understand both the nature of the problem and the structure of the search space. The various merits of using EC methods to address the CPP are discussed in the following section. 5.2.1 An overview of the Greenfield CPP The copper network provides customer access to a range of narrow and mid-band services. Most copper architectures comply with a three tier model consisting of the primary, secondary and distribution networks, all implemented in a tree structure. The primary network connects the secondary network to the exchange via a Primary Connection Point (PCP), as shown in Figure 5.1. The distribution network connects customers to Distribution Points (DPs), which are connected via the secondary network to the PCP. Access networks are normally accommodated within an underground duct network. The duct network forms the links between the access nodes, such as joint boxes, which are used to accommodate cable joints and DPs. The duct network is normally highly interconnected to provide flexible schemes routing between the exchange and customer. A typical secondary network structure connecting customers to PCPs, is shown in Figure 5.2, In this context, local access network planning is defined in the following manner. For a predefined geographical area, duct structure and particular demand profile determine the following: Telecommunications Optimization: Heuristic and Adaptive Techniques82 DP DP DP DP PCP PCP Main Pairs Cables Direct Pairs cables Primary Network Secondary Network Distribution Network = Customer DP Pairs DP = Distribution Point = Primary Connection Point Exchange Figure 5.1 Copper access network. 1. location of all access nodes (footway boxes), 2. location of DPs, 3. assignment of customers to DPs, 4. aggregation of DPs into sub-networks, 5. assignment of DPs to cable paths, and 6. route of all cables to satisfy customer demand at the lowest cost. The distribution network architecture is built from a number of different types of sleeves, joint boxes, cables and ducts. Customers are connected to the network at connection points called sleeves, which are normally accommodated in underground joint boxes. Ducts are used to house the secondary and distribution multicore cables, which range in size from 5 to 100 pairs. 5.2.2 Network Object Model A model of the network was created using object-oriented development methods, Object oriented analysis OOA and design OOD enable development of flexible, intuitive models. Object-oriented models are based upon the underlying domain structure. A system built around domain entity structures is normally more stable than one based around functionality. It has been recognised that most software modifications are associated with system functionality rather than the underlying domain (Paul et al., 1994). Addressing Optimization Issues in Network Planning with Evolutionary Computation 83 PCP Sleeve DP Customer Cable Duct Figure 5.2 Secondary copper network. These principles are then applied to create a model (Paul et al., 1994). Classification (abstraction) is the operation of grouping together objects with similar features. Inheritance (generalisation) if there are specialised types of a class which exhibit common features, the principle of inheritance can be used to organise them. Association represents a relationship between two different classes of objects. Aggregation (whole/part structure) expresses a special relationship in which one class contains another. As object modelling is based upon the organisation principles of human thought processes, domain specialists are able to contribute to the design of an object model without requiring specialist computer skills. In an object-oriented model, the objects and their associations correspond directly to real world entities and their relationships. Consequently, the object model is modular and closely resembles the real network. Objects can represent both physical items (copper cables) and abstract concepts (time slots in a frame). These characteristics make object models easy to understand, maintain and extend. A network object model is required to capture the network topology and its connection rules. Persistent computer based objects capture the functionality of network components and state attributes. The use of object-oriented methods proved successful because they allowed the complex rules relating to network component connectivity to be represented in a flexible and concise manner. This scheme allows the seamless integration of engineering rules into the computer based objects. The majority of constraint checking for engineering rules is therefore carried out implicitly, through the mechanisms provided by the model. Telecommunications Optimization: Heuristic and Adaptive Techniques84 However, practical experience has shown that another level of data abstraction is required to improve optimisation efficiency. The direct manipulation (creation and deletion) of network objects by the solver module during an optimisation run proved to be highly inefficient. To improve the efficiency of the process, relevant data from the object model is now held in the form of tables that can be accessed very rapidly. The design of the object schema employed the Object Modelling Technology (OMT) method developed by Rambaugh. This object model uses the class Nodes to represent access points, equipment cabinets and network components and the class Links to represent the cables and ducts. The high level view of the class schema is shown in Figure 5.3. The general-purpose object model has also been employed to support network performance and reliability analysis. Figure 5.3 General telecommunications network object model. 5.2.3 Structural Analysis Module The function of this module is to analyse network structure before an optimisation run. This pre-process generates data held in a distance matrix, tabulating the distance and optimal route between any two nodes. Graph theory is used to calculate the shortest distance between any two nodes and form a modified minimum spanning tree. Floyd’s Algorithm is applied to generate the distance and path matrices. Distance information is required to initialise the search algorithm. 5.3 Copper Planning Tool An integrated network planning environment based upon a Windows Graphical User Interface (GUI) has been created. This environment provides interfaces to existing external databases so that planning data can be easily transferred via a local area network. Addressing Optimization Issues in Network Planning with Evolutionary Computation 85 The GUI allows the user to view networks in a number of different ways. A network is normally viewed against a geographical map background. The display is layered so that alternative layers can be switched on and off as required. A colour scheme and a standard set of icons have also been created so that the components shown on different layers can be readily identified. A pan and zoom feature is also available to facilitate the inspection of complex network structures. In addition, network designs can also be shown in a simplified schematic format. The copper optimisation planning system has three modes of operation: manual, partially automated and fully automated. The manual option allows the planner to specify the placement of cables and duct components and to identify the location of DPs at network nodes. By using the partially automated option, the system determines the DP layout scheme and locations where cables can be inserted into the network. The fully automated option further increases the copper planning productivity, selects the network component types, provides routing information for ducting and cabling, optimises the cost and fully determines the connection arrangement of sleeves. A high level schematic diagram of the planning system is in Figure 5.4. The core of the fully automated option involves the application of evolutionary computing methods and graph theory to produce optimal or near-optimal solutions. A map representing the area under consideration is digitised and sent to the input of the system. This scanned map provides the Ordinance Survey (OS) co-ordinates needed to accurately locate network component nodes and customers’ premises. A network of possible duct routes along trenches is shown on the civil layer and overlaid on the scanned OS map. The subsequent optimisation process uses the duct network data. Users may interactively modify the network data stored in memory to customise network designs. Input Files Graphic Display Object Oriented Data Model 2 Genetic Algorithms & Graph Theory Post Processing Ducting Information Cabling Information Processing Options Semi- Auto Fully- Auto Manual Figure 5.4 Copper planning system overview. The objective of the fully automated option is to determine the best locations for DPs and form geographically optimal sub-networks. The duct network specified by the input data file for the optimiser can be configured as either tree or mesh networks. A tree Telecommunications Optimization: Heuristic and Adaptive Techniques86 structure is required for each solution within a sub-network to aggregate cables from customers to DPs and DPs to a PCP. The whole process is performed rapidly and cost- effectively without violation of technical constraints. Figure 5.5 shows a network represented on the civil layer. Figure 5.6 shows DP locations and tree formations within each sub-network identified by the optimiser. Cable joints are not required within the nodes shown as dotted circles. 5.4 Problem Formulation for Network Optimisation A specific planning problem is defined by an access network structure in terms of nodes and links and by the customer demand associated with the termination nodes. The principal aim of access network expansion planning is to determine the component and cable layouts for a given network, so that all customer demand is satisfied, at the minimum cost. A solution to the problem details information relating to all network components and cables. Demand may be satisfied by single or multiple sub-networks. More formally, an optimisation problem is defined by a set of decision variables, an objective function to be minimised and a set of constraints associated with the decision variables. The set of all decision variables defines a solution to the problem. Decision variables are the location, quantity, type and connection arrangements for all copper system components. The objective function, incorporating the decision variables, allows the evaluation and comparison of alternative solutions. In this case, the objective function determines the actual cost of copper access network installations. There are two types of constraints on solutions. The first set of constraints ensures that all customer demand is satisfied by a solution, whereas the second set of constraints ensures the technical feasibility of a solution. Expansion planning can be visualised as a mapping process in which one or more sub- networks are mapped onto a given access network duct structure, as depicted in Figure 5.6. During this process, components must be assigned to access nodes and cables to ducts. The total number of different mappings to evaluate is enormous. This is due to the vast range of possible configurations arising from alternative component combinations and the potentially large number of network nodes. Access network expansion belongs to the class of constrained non-linear combinatorial optimisation problems. It has been recognised that network planning problems are generally difficult to solve because of the size and structure of the search space. The size of the search space for an access network expansion problem of moderate dimensions is already so large that even the fastest currently available computer would need far longer than the estimated age of the universe to search it exhaustively. The objective function and imposed constraints of the optimisation problem are non-linear and the decision variables have a discrete range. In addition, the problem cannot easily be converted into a form where existing solution procedures can be applied and guaranteed to succeed. All of these factors, considered as a whole, make the successful application of standard optimisation methods very difficult to achieve. In some cases, conventional solution methods cannot be applied successfully. Furthermore, in such situations, a near optimal solution is the best that can be expected and consequently heuristic methods are often the only viable option. Addressing Optimization Issues in Network Planning with Evolutionary Computation 87 Figure 5.5 Network showing a set of potential DP locations. = Customer = Primary Connection Point Subnetwork 1 Subnetwork 2 = Distribution Point Figure 5.6 Optimised network showing DP locations and tree formation in each sub-network. Telecommunications Optimization: Heuristic and Adaptive Techniques88 A realistic formulation of the problem at the outset is critically important because it directly affects the chances of success. If the problem formulation is oversimplified then solutions will be invalid because they do not address the full or real problem. However, if too many factors are taken into account the problem can be made over complex, to the extent that it may prove to be impossible to comprehend and solve. These conflicting factors must be considered during the algorithm design phase. 5.5 Evolutionary Algorithms Evolutionary Algorithms (EAs), introduced in chapter one, are part of a family of heuristic search algorithms (Holland, 1975; Goldberg, 1989). EAs have already been used successfully to solve other difficult combinatorial optimisation problems. An important consideration is that EAs do not require special features such as linear or differentiable objective functions and are therefore applicable to problems with arbitrary structures. Evolutionary methods seek to imitate the biological phenomenon of evolutionary reproduction. The principles of evolutionary computing are based upon Darwin’s theory of evolution, where a population of individuals, in this case potential solutions, compete with one another over successive generations, as in ‘survival of the fittest’. After a number of generations the best solutions survive and the less fit are gradually eliminated from the population. Instead of working with a single solution, as is the case for most optimisation methods, EAs effectively manipulate a population of solutions in parallel. As solutions evolve during an optimisation run, the mechanism ‘survival of the fittest’ is employed to determine which solutions to retain in the population and which to discard. Each solution is evaluated and promoted into the next generation according to its fitness for purpose. The selection is performed stochastically and the probability of a solution surviving by moving into the next generation is proportional to its fitness. This process allows less fit solutions to be selected occasionally, and therefore helps to maintain genetic diversity in the population. In an EA, the decision variables are normally represented in the form of an encoded string. The individual variables are called genes and the sum of all genes is called a chromosome. A fitness value for each solution is assigned according to a fitness function. In this particular case, the fitness function evaluator module is closely associated with the object model. The Genetic Algorithm (GA) is a particular type of EC algorithm. In a GA, new solutions are produced after the application of genetic operators. For a canonical GA the fundamental operators are crossover and mutation. Crossover is based upon the principle of genetic mixing during which genes from one chromosome are mixed with those of another. Mutation, on the other hand, is an operator which works on a single chromosome, normally only making small changes to the genetic information. Mutation contributes to the maintenance of genetic diversity, which in turn helps to avoid premature convergence. Recently, EC methods have attracted a considerable amount of attention because they have been successfully applied to solve a wide range of complex and difficult problems from various disciplines, including telecommunications, logistics, finance, planning, transportation and production. [...]... encoding, fitness calculation, selection, crossover, mutation and replacement operators Standard crossover, mutation and replacement strategies are provided by default 90 Telecommunications Optimization: Heuristic and Adaptive Techniques Alternative string encoding schemes and problem specific GA operators can be added by over-riding the existing functions The GA class also allows the programmer to create... number of sub-networks required by considering the total demand of the customers 2 Define the encoded string length, using the estimated number of sub-networks required 92 Telecommunications Optimization: Heuristic and Adaptive Techniques 3 Encode an integer string with elements representing the centre of each sub-network 4 Create an initial population by randomly generating a set of integers between 1... Addressing Optimization Issues in Network Planning with Evolutionary Computation Figure 5.9 Copper planning tool GUI Figure 5.10 A complete layout for a practical network 93 94 Telecommunications Optimization: Heuristic and Adaptive Techniques Sub-Net 2 Demand = 98 Sub-Net 1 Demand = 70 Figure 5.11 Optimised solution showing DP locations, with demand distribution bias factor disabled In this case, the total... in a single cabinet, (ii) the cost of travelling and setup time These factors have been shown to have a significant effect upon the structure of network solutions 96 Telecommunications Optimization: Heuristic and Adaptive Techniques • Support for cable ‘back feeding’ is provided In some cases, it is necessary that cables connected to customers follow a duct path directed away from the exchange • The . data Telecommunications Optimization: Heuristic and Adaptive Techniques, edited by D. Corne, M.J. Oates and G.D. Smith © 2000 John Wiley & Sons, Ltd Telecommunications Optimization: Heuristic and Adaptive. structures. Macro-level network system analysis can now be attempted because of developments in heuristic, mathematical and intelligent methods, and the availability of low cost high performance. Ltd ISBNs: 0-471-98855-3 (Hardback); 0-470-84163X (Electronic) Telecommunications Optimization: Heuristic and Adaptive Techniques80 using new smart algorithms, thereby providing better solutions