**6. Conclusion**

20 Will-be-set-by-IN-TECH

tile size can be a very valuable input of the neural network to correctly classify the cachability

During the training process, a training record corresponding to the request of a particular tile is associated with a boolean target (0 or 1) which indicates whether the same tile is requested

Once trained, the neural network output will be a real value in the range [0,1] that must be interpreted as the probability of receiving a successive request of the same tile within the time window. A request is classified as *cacheable* if the output of the neural network is above 0.5.

The neural network is trained through supervised learning using the data sets from the extracted trace files. The trace data is subdivided into training, validation, and test sets, with the 70%, 15% and 15% of the total requests, respectivelly. The first one is used for training the neural network. The second one is used to validate that the network is generalizing correctly and to identify overfitting. The final one is used as a completely independent test of network

Each training record consists of an input vector of recency, frequency and size values, and the known target. The weights are adjusted using the backpropagation algorithm, which employs the gradient descent to attempt to minimize the squared error between the network output values and the target values for these outputs [36]. The network is trained in batch mode, in which weights and biases are only updated after all the inputs and targets are presented. The

Neural network performance is measured by the correct classification ratio (CCR), which computes the percentage of correctly classified requests versus the total number of processed

CartoCiudad IDEE-Base

pocket algorithm, which saves the best weights found in the validation set, is used.

training 76.5952 75.6529 validation 70.2000 77.5333 test 72.7422 82.7867

**Table 5.** Correct classification ratios (%) during training, validation and testing for Cartociudad and

Figure 16 shows the CCRs obtained during training, validation and test phases for Cartociudad and IDEE-Base services. As can be seen, the neural network is able to correctly classify the cachability of requests, with CCR values over the testing data set ranging between 72% and 97%, as shown in Table 5. The network is stabilized to an acceptable CCR within 100

<sup>1</sup> *if the tile is requested again in window*

<sup>0</sup> *otherwise* (5)

of requests.

generalization.

requests.

IDEE-Base.

to 500 epochs.

again or not in window, as shown in Equation 5.

Otherwise, it is classified as *non cacheable*.

*target* =

Serving pre-generated map image tiles from a server-side cache has become a popular way of distributing map imagery on the Web. However, in order to achieve an optimal delivery of online mapping, adequate cache management strategies are needed. These strategies can benefit of the intrinsic spatial nature of map tiles to improve its performance. During the startup of the service, or when the cartography is updated, the cache is temporarily empty and users experiment a poor Quality of Service. In this chapter, a seeding algorithm that populates the cache based on the history of previous accesses has been proposed. The seeder should automatically cache tiles until an acceptable QoS is achieved. Then, tiles could be cached on-demand when they are first requested. This can be improved with short-term prefetching; anticipating the following tiles that will be requested after a particular request can improve users' experience. The metatiling approach presented here requests, for a given tile request, a bigger map image containing adjacent tiles, to the remote WMS backend. Since the user is expected to pan continuously over the map, those tiles are likely to be requested. Finally, when the tile cache runs out of space, it is necessary to determine which tiles should be replaced by the new ones. A cache replacement algorithm based on neural networks has been presented. It tries to estimate the probability of a tile request occurring before a certain period of time, based on the following properties of tile requests: recency of reference, frequency of reference, and size of the referenced tile. Those tiles that are not likely to be requested shortly are considered as good candidates for replacement.

#### **Acknowledgements**

This work has been partially supported by the Spanish Ministry of Science and Innovation through the project "España Virtual" (ref. CENIT 2008-1030), a FPI research fellowship from the University of Valladolid (Spain), the National Centre for Geographic Information (CNIG) and the National Geographic Institute of Spain (IGN).
