Evolutionary Computation
GA 1 Select Customer Distribution Point (DP) Steps 1 to 5 are only required during initialisation
1. Employ Floyd’s algorithm to create distance and path matrices to describe network connectivity.
2. Specify the population size and number of generation.
3. Define the encoded string length, using the total number of DPs.
4. Encode a binary string to represent selected DPs.
5. Create an initial population by randomly generating 0s and 1s.
6. Form a tree network to connect customers, select both DPs and Primary Connection Points (PCPs).
7. Evaluate the fitness of each string using the cable distances and equipment costs.
8. Define a sub-population for mating using roulette wheel selection.
9. Apply crossover to the binary encoded strings.
10. Perform mutation according to the mutation probability.
11. Replace the old population with the new population.
12. Repeat steps 6 to 12 until the number of generations reaches a predefined terminal value.
GA2 – Select Sub-Network DP Cluster
Steps 1 to 4 are only required during initialisation.
1. Estimate the 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.
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 and the number of selected DPs determined by the first DP selection GA.
5. Evaluate the fitness of each string by considering the equipment cost and the cable distance of each customer from the nearest DP specified in the string.
6. Impose a penalty cost if the total demand of a sub-network violates the maximum allowable demand and consider the required distribution of spare capacity.
7. Define a population for mating using roulette wheel selection.
8. Apply crossover on the (integer) encoded string.
9. Perform mutation according to the mutation probability.
10. Replace the old population with the new population.
11. Repeat steps 5 to 11 until the number of generations reaches a predefined value.
The tool also allows the user to enter and manipulate data via the graphical interface.
Data is automatically validated and checked within the system before being passed to the optimiser. An important advantage of this tool is that it produces accurate and consistent computer based network records.
5.6 Practical Results
A section of a real solution for a practical network is shown in the smart copper planning tool environment (Figure 5.9). The complete layout of this practical example and the optimised solution can be found in Figures 5.10 and 5.11, respectively. In this case, the tool provided a satisfactory solution with acceptable DP distributions and customer clusters. A total time of less than 13 minutes was taken to solve this 240-node problem. This example clearly demonstrates the effectiveness of EC methods when used to solve practical network problems. Table 5.1 gives a summary of the results and algorithm settings.
A special feature of the GenOSys tool is the ability to control the degree of spare capacity distribution. The distribution of the demand capacity for each sub-network can be adjusted by a special optimiser control variable, the demand distribution bias factor.
The optimised solution, given in Figure 5.11, shows the DP layout without a penalty imposed to control the spare capacity distribution. The total service demand for sub-networks 1 and 2 is 70 and 98, respectively.
Table 5.1 Summary of results and EA settings.
Network results Number of nodes; a network of realistic size 240
Number of runs 20
Number of generations 3000
Population size 50
Average elapsed time in minutes for each run (in P166 machine) 12 Average deviation from the best known minimum cost < 1%
Figure 5.9 Copper planning tool GUI.
Figure 5.10 A complete layout for a practical network.
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 service demand for both networks is 168. As each sub-network can only accommodate 100 cable pairs, at least two sub-networks will be required. The criterion applied to form optimal sub-networks is based solely on the shortest distance between customers to DPs and DPs to a PCP. In this example, the demand distribution for each sub-network may be considered to be out of balance, with demand levels of 70 and 98. This design dictates that there is a relatively large spare capacity for the first network, but only a small amount for the second.
Figure 5.12 shows what could possibly be a more desirable DP layout, exhibiting a more even distribution of capacity. In this modified design, the service demand for sub-networks 1 and 2 becomes 83 and 85. A consequence of this design strategy is that the modified network will require longer cables and be marginally more costly to implement. In this mode of operation, the optimiser employs a MOGA to plan (i) a low cost network, which (ii) also exhibits an evenly distributed capacity between sub-networks.
5.7 General Discussion
As discussed, network planning problems are inherently difficult to solve. An added complication and an important factor to consider is the customer demand forecast level. It is
sometimes difficult to apply forecasting methods to derive an accurate and reliable forecast level for planning purposes. Unfortunately, there is inevitably a degree of uncertainty associated with demand level forecasts.
Sub-Net 2 Demand = 85
Sub-Net 1 Demand = 83
Figure 5.12 Optimised solution showing DP locations with demand distribution bias factor enabled.
The planner can control the level of a spare capacity in a sub-network by adjusting the demand distribution variable. The purpose of this feature is to aid the management of networks that are characterised by uncertain data. Smart computer based tools rapidly generate solutions, enabling the planner to select between alternative network configurations after interactively changing the design criteria.
To summarise, GenOSys has the following principal features:
• GenOSys is able to process networks defined as mesh or tree network structures. Graph theory (Floyd’s Algorithm) is used to analyse network structures.
• A highly detailed cost model has been created which takes into account, (i) the co-location of equipment 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.
• 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 optimiser uses a hybrid genetic algorithm and is intelligently seeded so that only valid solutions are added to the initial population.
• The system has been developed using Object-Oriented Methods (OMT) and a generic toolkit of software components for optimisations has been created.
• Rapid Application Development (RAD) techniques have been used throughout the development, using the programming language Delphi.
• The tool is easy to set up because users need only define the demand distribution bias factor to ensure an even distribution of spare capacity across the network.
• A cable aggregation function is built into the optimiser. An important feature is the ability to design optimal layout schemes for multicore cables.
• Very large and complex networks can be solved rapidly, with more than 500-nodes.
• Planners are now able to experiment and adapt network solutions using sensitivity analysis techniques.
• A MOGA allows the planner to customise network layouts to satisfy specific design criteria.
As this tool has been developed in collaboration with BT the dominant components of the CPP have been retained in the formulation.
5.8 Benefits for Management
GenOSys has been used throughout BT to plan greenfield sites. This tool has now been in use for about a year and reports from users relating to its performance have been very positive. It is evident that productivity levels have improved dramatically, between twenty and hundred fold. Concurrent working methods have also been successfully applied and work backlogs are gradually being removed. In the future, it is anticipated that fewer offices will be needed to process all new BT UK copper planning projects.
Furthermore, Asumu and Mellis (1998) report that, “automatic network optimisation offers an additional improvement in guaranteeing the cost optimal solution (typically 20%
better than the average planner) but the main benefits of automatic planning are greater standardisation, consistency and quality, even with less skilled operators”. The GenOSys tool is a component within a suite of smart tools under development in BT. They also stated that, “if the key requirements of modularity and data-sharing are adhered to, the results will be a cohesive set of network planning tools, offering massive planning productivity and improvements, the better integration of network planning and management, and the guarantee of optimally designed, high performance networks”. Subsequently, they reported major savings in capital expenditure. Finally, 'the set of tools described here have been successfully proven in test-bed trials and the business benefits projected to accrue from their field deployment is about £10 million per annum, or about 10% of the total annual investment made in the copper greenfield network' (Asumu and Mellis, 1998).
For users, the tasks associated with network planning have changed and now mostly involve data preparation and creative interactive design. Computer based interactive design practices have largely eliminated inefficient manual methods. However, a great deal of work remains to be completed because current network records are usually held on paper drawings. The conversion and transfer of all paper records into the digital data domain is now under way.
This work has shown that EA methods can be usefully employed in a practical setting to design telecommunications networks, by demonstrating that they can rapidly produce solutions exhibiting near ideal low cost structures.
Smart planning tools allow designs to be produced more rapidly with automatically created component inventories and orders may be processed via electronic interfaces between suppliers and contractors. These advances make the application Japanese manufacturing methods in the telecommunications sector a real option. Just-in-time manufacturing methods are have the advantage of reducing response times, stock inventory levels, overheads and capital expenditure.
5.9 Strategic Issues
The underlying objective of UK Government policy is to provide customers with the widest range of telecommunications services at competitive prices. Current Government policy has concentrated on creating a framework for infrastructure development and investment. As the process of European liberalisation continues the level of competition is gradually increasing in the telecommunications sector. It is now essential that network operators have the necessary analytical, financial and customer oriented skills to meet the challenge of providing a wide range of reliable, low cost services.
In the telecommunications sector, sustainable competitive advantage can be achieved by applying new ‘intelligent’ technologies which empower managers, enabling them to analyse complex systems, determine appropriate management strategies and expedite projects and business cases. Implementation of the methods described in this section provides managers with new smart tools to manage complexity and change, thereby providing improved insight into, and control over, network operational issues.
Acknowledgements
This research has been conducted in collaboration with BT Labs. The authors would like to thank BT staff, Don Asumu, Stephen Brewis and John Mellis, for their technical input and support. In addition, the authors would also like to thank the following PhD researchers, Harald Paul and Christian Woeste, for their valuable work on the prototype PON optimisers.