GB2386287A

GB2386287A - A method of monitoring state of a telecommunications network

Info

Publication number: GB2386287A
Application number: GB0214520A
Authority: GB
Inventors: Lester Tse Wee Ho; Jonathan Michael Pitts; Louis Gwyn Samuel
Original assignee: Lucent Technologies Inc
Current assignee: Nokia of America Corp
Priority date: 2002-03-06
Filing date: 2002-06-24
Publication date: 2003-09-10
Anticipated expiration: 2022-06-24
Also published as: GB2386287B; GB0214520D0; GB0205284D0

Abstract

The present invention is to a method of monitoring state of a telecommunications network, and to a telecommunications network comprising a complexity determinator. The network comprises a plurality of nodes, each node is at a time on one of a number of node states dependent upon a predetermined set of rules and local conditions at the node. The method/ complexity determinator comprises determining the value of a parameter dependent upon entropy (which is a measure of how random the system behaviour is) of the node states in the network. The present invention provides entropy based complexity measurement for behaviour analysis of self organising telecommunications networks. Specifically, the application of the entropy-based complexity measure in telecommunications networks is useful to examine behaviours of localised distributed algorithms (LDA), (operative at Base Station levels and uses only local information and simple local rules at the Base Stations). The behaviour of self-organising LDA networks can thus be investigated to determine network-wide trends. The complexity measurement is used for setting the size of the Base Station cells.

Description

A METHOD OF MONITORING STATE OF A TELECOMMUNICATIONS NETWORK COMPRISING A PLURALITY OF NODES, AND A CORRESPONDING TELECOMMUNICATIONS NETWORK Technical Field The present invention relates to a method of monitoring state of a telecommunications network comprising a plurality of nodes. The present invention also relates to a telecommunications network comprising a complexity determinator and plurality of nodes.

Background of the Invention Mobile telecommunications networks are mostly now centrallycontrolled, but are evolving towards more ad-hoc and dynamic structures with cheap, low-powered nodes, such as base stations, that are auto-configurable and flexible.

Controlling such networks means coping with uncertainty, not only in traffic demand but also in the structure of the network itself. Using decentralised control gives flexibility to respond locally to these uncertainties, creating self-organising networks that rely on time-varying i. e. emergent behaviour globally across the network so as to provide network-wide co-ordination. Compared with more easily monitored centrallycontrolled networks, however, these decentralised networks can be difficult to predict or manipulate.

Mobile networks are expected to be used less for voice-oriented services and more for data services. As the anticipated demand for data oriented applications increases, so-called 3G systems (i. e those in accordance with Third Generation Partnership Project 3GPP standards) are expected to meet the demand as thy evolve into future generation, or fourth generation (4G) networks. In order to accommodate the larger amount of traffic, 4G networks are expected to use a higher frequency band and offer channels that have bit rates that are ten times higher than that of 3G systems, as mentioned in T. Otsu, 1. Okajima, N. Umeda, Y. Yamao,"Network architecture for mobile communications systems beyond IMT-2000", IEEE Personal Communications,

volume 8, number 5, pages 31-37, October 2001. The high data transmission rates that are required will necessitate the use of smaller cell sizes, and hence, potentially cause a serious increase in the cost of the network infrastructure and the cost of planning and deploying the network, as outlined in Y. Yamao, H. Suda, N. Umeda, N. Nakajima.

"Radio access network design concept for the fourth generation mobile communication system", IEEE 51, t Vehicular Technology Conference Proceedings, 2000, VTC 2000- Spring Tokyo, volume 3, pages 2285-2289. This, among other reasons, has prompted suggestions to use cheap, low-powered devices that are highly flexible and are able to auto-configure to produce networks that are more ad-hoc and dynamic in structure. as described in various papers such as: A. Bira, F. Gessler, O. Queseth, R. Stridh, M.

Unbehaun, J. Wu, J. Zander, "4th-Generation wireless infrastructures : scenarios and research challenges", IEEE Personal Communications, volume 8, number 6, pages 25- 31, December 2001, in J-Z. Sun, J. Sauvola, D. Howie,"Features in future: 4G visions from a technical perspective", Global Telecommunications Conference, 2001.

GLOBECOM'01, volume 6, pages 3533-3537, in B. G. Evans, K. Baughan,"Visions of 4G", Electronics and Communications Journal, volume 12, number 6, pages 293-303, December 2000, and in Wireless World Research Forum,"Book of Visions 2001 : Visions of the Wireless World", http://www. wireless-worldresearch. org/BoVl. 0/BoV/BoV2001vl. O. pdf.

Decentralised control offers the flexibility and robustness needed to cope with the dynamic nature of such ad-hoc networks, and has many advantages over centralised control, as discussed in B. Xu, B. Walke,"Design issues of self-organizing broadband wireless networks", Computer Networks: the International Journal of Distributed Informatique, volume 37, numberl, pages 73-81. Sept. 2001.

However for decentralised control, wireless networks would have to evolve to structures that have many similarities with high-speed wireless LANs. These networks would be built to carry all traffic through a common packet switch transport method (such as IP). Using high frequency bands to provide the required bandwidth. the maximum size of the cells would be a lot smaller than current systems because of the higher propagation loss.

Distributed (i. e decentralised) control of wireless networks is thus becoming increasingly important in the move towards self-configuring ad-hoc type networks for G (UMTS) and 4G networks. This ad-hoc approach to wireless networks

relies on highly distributed nodes to create flexible and highly robust network. These nodes typically handle or have the capability to handle the control of the network, which in a normal wireless network would have been done by a central controller. Functions such as routing, resource management, and auto-configuration are aspects that can be handled by ad-hoc networks. One of the challenges with an ad-hoc type network is that the amount of overhead signalling that has to be done to coordinate the whole network without a central controller can be impractical. especially if the size of the network is large. In consequence, distributed algorithms are used that make use of localized information, i. e. where the network nodes make decisions based on very limited information on the network, often only considering the information of a few neighbouring nodes. These localized algorithms often rely on the global behaviours that emerge out of the simple interactions between the network's nodes to produce the selforganization that keeps the networks in check.

One of the problems that arise from the use of such localized distributed algorithms (LDA) is that the interactions and behaviour of the network, although giving the desired effect, are not always straightforward. When using a centralized control approach, the behaviour of the network is always known and guided through algorithms contained in the controller. Since this is not available when using localized distributed algorithms (LDA), the behaviour of the network can be difficult to predict and debug if the behaviour starts to go out of hand, such as the synchronization problem found in the TCP/IP protocol, and the observation of criticality in self-organized systems. LDA's can also be very limited in their functionality. Often, the evolution of the behaviour of the network is one directional, where the network self-organizes towards one type of behaviour. which can limit the flexibility of the network.

Various techniques are known to be used to investigate the behaviour of self-organizing systems so that a better understanding on the effects of various parameters of an LDA. These methods are usually used in the physical sciences field to examine emergent phenomena. For example, Crutchfield, J. P.,''The Calculi of Emergence: Computation, Dynamics and Induction, " Physica D, voL75, no. 1-2, pp. 11- 54 describes use of state reduction to extract a state transition model of the system to represent its behaviour. The state reduction technique is quite complicated and requires a large sampling of the system behaviour before a sufficiently accurate state transitional model can be obtained. Of course, as the sample size is not infinitely large, we can

never determine for certain that the state transitional model that is produced is accurate enough.

Grasberger, P.,"Toward a quantitative theory of self-generated complexity,"International Journal of Theoretical Physics, vol. 25, no. 9, pp. 907-938.

Sept. 1986,), Lopez-Ruiz R. , Mancini H. L. , Calbet X.,"A statistical measure of complexity", Phvsical Letters A, pp. 321-326, 25 December 1995] and Shiner J. S. , Davison M., Landsberg P. T.,"Simple measure for complexity". Physical Review E, vol.

59, no. 2, pp. 1459-1464, February 1999 are papers which all describe use of entropy measurements of the system states to quantify the behaviour of the system in terms of state space. Entropy-based techniques capture the behaviour of the system, but also require a large sampling size of the system states to produce an accurate picture of the behaviour. This is because they consider the instantaneous state of the system, and not the overall trend of the behaviour, thus requiring a longer time to build up a sufficient behaviour representation.

Looking at the potential problems in distributed control of a wireless network, it was realised that changes needed to be made in the way current wireless networks are designed, deployed and maintained. The base stations have to be cheap, small and unobtrusive. To reduce the increased cost of deploying a larger number of cells, the base stations also need to have the ability to self-configure to a certain extent.

The process of installing a base station should be simple and straightforward, with a "plug and play"approach to installation, with the base station self-configuring different aspects of its operation such as its cell size and routing. It was realised that decentralised control offers this level of flexibility in the network and also helps with solving the scalability and robustness issues, and further decrease the cost of the network by removing the need for expensive equipment otherwise needed in centrally controlled networks. But highly decentralised networks work on self-organising behaviour, and can be unpredictable and difficult to manipulate.

For large networks composed of simple, inexpensive and highly decentralised nodes to co-ordinate themselves, they would have to be self-organising.

Self-organising systems have the ability to evolve and adapt to retain a certain coordinated behaviour using only localised information and relatively simple rules. It is a behaviour that relies on the emergent global behaviour arising from the interactions between the sub-systems. It is a behaviour that can be observed in many different

systems, ranging from physical, biological, sociological and mathematical systems. A main characteristic of self-organising systems is their ability to evolve towards the same behaviour pattern no matter what the initial configuration of the system is. Self- organising behaviour would be useful applied to co-ordination of a network. One can imagine a network where the base stations, using basic rules, would be able to settle down to the same behaviour regardless of what the initial conditions are during initial deployment, or what changes are made to the network.

While self-organisation has a very promising application in wireless networks, there are several problems and challenges that can arise from the use of selforganising networks. Even though self-organising networks are able to maintain a coordinating behaviour, under certain conditions, very sudden changes can occur. These sudden changes in the system behaviour, sometimes called self-organised criticality, occurs when certain parameters of the system are changed and changes the principle behaviour of the system, causing a phase transition, flipping it from, say, a static behaviour to a chaotic one as described in P. Bak, C. Tang, K. Wiesenfeld,"Self- organized criticality", Physical Review a (General Physics), volume 38, number 1, pp. 364-74, 1 July 1988. Examples of these critical behaviours have been observed in many different systems, including telecommunications networks as described in R. V.

Sole, S. Valverde,"Information transfer and phase transitions in a model of Internet traffic"Physica A, volume 289, number 3-4, pages 595-605,15 Jan. 2001. One of the features of criticality is that only a slight incremental change of a system parameter or structure can cause this change of behaviour. This characteristic is obviously not desirable in a wireless network where such a catastrophic failure could bring down the whole network very suddenly and without warning.

Another challenge that is posed by using self-organising networks is the difficulty of designing the algorithms themselves. Controlling the behaviour of centralised networks is more straightforward as the behaviour of the whole network is always known and can be changed directly. In self-organising networks, producing the algorithm that would give rise to the desired emergent behaviour for self-organisation requires more careful consideration, since producing the desired emergent behaviour cannot be done by simply telling the network nodes what to do directly.

The unpredictability of self-organising systems thus poses difficulties when applied to wireless networks. It would be advantageous to be able to determine

how the network would behave under different scenarios during the design of the network, for example to make sure that during the operation of the network there is some sufficient warning of unpredicted network behaviour. Also if something goes wrong, it would be useful to be able to readily analyse the network behaviour to find out the cause of the network failure.

To review. the use of self-organisation in for example 4G networks is important. Mobile networks are evolving from using a centrally controlled architecture towards one with a more ad-hoc and dynamic character. In such networks, the coordination and management system has to cope with a network hierarchy, connectivity and node availability that is not clear and changes continuously.

Distributed, decentralised controlled networks are ideal for use under these circumstances. Self-organising behaviour of these networks is capable of providing a very flexible and robust structure without using very complicated and expensive hardware and software, giving network providers the ability of providing high bandwidth services whilst keeping the cost of network deployment and management low. The difficulty of predicting the behaviour of self-organising networks, however, is an issue that can cause problems in decentralised networks. Self-organising systems are sometimes known to exhibit critical behaviour where, under certain circumstances, the behaviour of the network changes very drastically. The difficulty of predicting the behaviour also makes the design and optimisation of these networks awkward.

Summary of the Invention The present invention provides a method of monitoring state of a telecommunications network comprising a plurality of nodes, each node being at a time in one of a number of node states dependent upon a predetermined set of rules and local conditions at that node, the method comprising determining the value of a parameter dependent upon entropy of the node states in the network.

Preferred embodiments of the present invention thus provide entropy based complexity measurement for behaviour analysis of self-organizing telecommunication networks. Specifically, application of the entropy-based complexity measure in telecommunications networks is useful to examine behaviours of localized distributed algorithms. The behaviour of self-organising i. e. localised distributed

algorithm (LDA) networks can thus be investigated to determine network-wide trends.

Effects of parameters can be investigated that are not immediately apparent through other means of observation, and preferred embodiments can be useful guides for changing or tweaking parameters and localised distributed algorithms to optimise or produce a desired behaviour. Detection and pinpointing of causes of unexpected and undesired behaviours are further advantages. Analysis of telecommunications networks using a complexity metric that uses system state entropy gives an indication of the behaviour of the network. The metric provides a tool that provides the ability to examine the critical behaviour of the network when designing and optimising the network, and also to provide advance warning on network failure.

Understanding the behaviour of self-organising wireless networks can provide a lot of different benefits. For example, self-organising networks can store a library of many different simple algorithms that are activated under certain conditions to trigger different types of desired emergent behaviours, giving us even more flexibility in the control of the network.

Preferably the value of the parameter dependent upon entropy is determined dependent upon sufficient node states being determined as having been changed after filtering out of fluctuations having less than a predetermined duration.

Preferred embodiments thus provide a filtering technique to extract a system behaviour trend in a network.

Preferably the parameter is dependent upon entropy S normalised by the

maximum entropy possible S Preferably the parameter is complexity and is ma, x.

proportional to}-. maux maux

Preferably the local conditions at that node are the node states of at least one of the neighbouring nodes to that node.

Preferably the network is a network for mobile telecommunications and the nodes are base stations. Preferably each base station operates according to a set of rules for automatic cell sizing defined in a local algorithm and takes at any one time one of three states in respect thereof. Preferably the three states are: cell sizing possible but not occurring, cell sizing occurring, and cell sizing not permitted.

The present invention also provides a corresponding telecommunications network. In particular, the present invention provides a telecommunications network comprising a complexity determinator and plurality of nodes, each node being at a time in one of a number of node states dependent upon a predetermined set of rules and local conditions at that node, the complexity determinator being operative to determine the value of a parameter dependent upon entropy of the node states in the network.

Preferably the determinator comprises a filter operative to filter out fluctuations having less than a predetermined duration, the determinator being operative to determine the value of the parameter dependent upon entropy dependent upon sufficient node states being determined as having been changed after filtering out of the fluctuations having less than the predetermined duration.

Brief Description of the Drawings A preferred embodiment of the present invention will now be described by way of example and with reference to the drawings, in which: Figure 1 is a diagram illustrating state transitions of a base station Figure 2 is an illustration of the state space matrix of the network, M at a time instant, Figure 3 shows complexity of network behaviour vs. back off time for a simulated network.

Figure 4 shows on-line complexity measurement graphs and the corresponding cell boundary diagrams, with back-off time set to (a) 15 seconds, (b) 20 seconds, and (c) 25 seconds, and Figure 5 shows (a) On-line complexity measure values during operation of the simulated network, back off time set to 25 seconds, one iteration being 0.6 seconds, and further base stations being added at 650 iterations ; and shows (b) the resulting cell configuration with the further base stations.

Detailed Description The inventors addressed the problems and difficulties of controlling decentralised networks by introducing a method of analysing and handling the behaviour of decentralised wireless networks using an entropy based complexity metric

An entropy-based complexity metric is used to investigate the behaviour of self-organising systems in mobile networks. Complexity in this context means the minimal state representation of the system behaviour. A complexity measure is applied to extract information on network-wide behaviour, the particular example behaviour being the self-organising behaviour involved in cell dimensioning. The metric is general enough for application in monitoring other behaviours. such as decentralised channel allocation behaviours, and is also applicable in other types of telecommunications networks.

To reiterate, a method is described below of monitoring the behaviour of networks operating with self-organizing, highly distributed, localized algorithms using complexity metrics. The method uses entropy measurements of the network states to track the behaviour of the network (e. g. chaotic, static, oscillating, or complex behaviours). The sampling of the system state is processed through a filtering mechanism so that the slight fluctuations in the network behaviour are not recorded and only the global trend in the behaviour is considered. Details are presented below, including an example of complexity measurement on an automated algorithm for setting the sizes of base station cells.

Complexity The term complexity is often used in the study of large, distributed systems. A system is here considered as a collection of interacting, interconnecting subsystems, where each subsystem has a number of finite states. The complexity of the behaviour of a system is defined in this context as the minimal representation of the behaviour of the system. In other words, a system that has behaviour that is simple to describe would have less complex behaviour than that with a behaviour that is more difficult to describe. Under this definition, a system that has a totally random behaviour would have no complex behaviour, as the behaviour can be described easily as white noise. Totally static systems also posses no complex behaviour, as its behaviour is obviously also very easy to describe. More complex behaviours are those that lie between these two extremes. They do not have a very random nature. nor are they very static. They have a dynamic but correlated behaviour, with persistent"structures"in the system state space.

Entropy Knowing the complexity of the behaviour of the system would give us an insight into its behaviour. We would be able to detect any changes in the basic behaviour of the system as a whole, using a very general method that can be applied to many different systems. Various efforts at quantifying the complexity of the behaviour have been made, with different approaches adopted to obtain the size of the representation of the system behaviour. Of the several measures developed. one of the approaches that is taken is by deriving the complexity measure using the entropy S of the system:

where p, is the probability that the system is in state i of N possible states.

The entropy of a system is a measure of how random the system behaviour is. It takes into account how often a system visits different areas of the system state space, and if the system visits all the states with equal probability (i. e. totally random), then the entropy will be at maximum. Conversely, if the system only visits one state (i. e. totally static), the system entropy will be zero. The entropy will give us an indication of the size of the system state space, and this has a relationship with the size of the representation of the system behaviour, as we know that maximum entropy gives us a perfectly random system, while zero entropy gives us a static system.

A complexity measure based on entropy We introduce a complexity metric that is derived from the system entropy. The complexity C, is defined as:

where S is the entropy of the system. and Smax is the maximum possible

entropy of the system that occurs when all states are equiprobable, i. e. when P, =-Vi N

and is given by logl o N. is basically the measured entropy normalised to the /S ma\

maximum entropy possible, and gives us a measure of the randomness of the system.

This complexity metric indicates low complexity at both random and static behaviours.

Example application in a wireless network having decentralised control of cell sizes By way of example, we apply the complexity measurements to a simple distributed algorithm for automatic cell dimensioning in a decentralised wireless telecommunications network. The algorithm is distributed in being operative at base station level and uses only local information and simple local rules at each base station.

A base station increases its cell size slowly in an attempt to poll its neighbours. When a neighbour is able to detect a polling signal, it sends an acknowledgement to the originating cell. Once the polling cell reaches its maximum cell size, it refers to the list of acknowledgements it has received and adjusts its cell size accordingly (for example, to take a cell radius of just over half the distance to the nearest neighbouring base station). If a base station is polled, then it will be triggered into the polling state as well, to reconfigure itself after a new neighbour has been switched on. The base stations are totally independent and autonomous, using no direct input from a central controller.

As shown in Figure 1, each base station is in one of three possible states at any one time: either polling, in a frozen buffer state or in an idle state. After completing the polling of its neighbours, the base station configures its cell size (dependent upon the distance to its neighbour (s), for example such that the cell radius is set to be just over half the distance to the nearest neighbouring base station) and enters the frozen buffer state for a fixed amount of time (called the back off time). This frozen buffer state is needed to prevent other neighbouring base stations that have been powered on at the same time from triggering a newly configured base station from entering the polling state again, resulting in a situation where the base stations are triggering each other off without a chance of completing their polling state. During the frozen state. the base station cannot be triggered into entering a polling state by another base station. At an idle state, the base station has successfully configured itself and is receptive to being triggered to the polling state by a neighbour.

The 100 base stations in the network model that was used in the simulations are loosely arranged in a 10 by 10 grid, each placed at approximately 200

metre intervals and are powered on at randomly set times. The base stations are recorded as having one of three states: polling, idle or frozen.

With three states and 100 base stations, there are, or approximately 1047 system states. Because of this huge amount of possible states, a state definition filter as described below is used to capture the basic trend of the network behaviour, rather than the"raw"states.

The network state space can be defined by a matrix. 1\1, where an element

k which has a value of either 1 or 0, shows whether the network node i is in state k (whereupon the value ouf ink is set to 1) or not (when lvi, is set to 0). Since we use three node states (idle. polling and active), and 100 nodes, Mis a 100x3 matrix. When recording the states of the network, we record, at each time step, a snapshot of the network state in the form of M, and pass it through a state definition filter to remove the slight fluctuations in the state space. Figure 2 is an illustration of a snapshot of the state space matrix of the network, M. A white box indicates which of the three states a base station is currently in.

Filtering Applying the entropy-based measure to the decentralised wireless network, several characteristics can be extracted from the results. When doing this, however, we use a state definition filter that removes any slight (i. e non-persistent) fluctuations in the system state when we are recording the system behaviour. The state definition filter causes only persistent behaviour to be recorded so that it is the basic overall behaviour of the system which is recorded. This is because we are interested only in the basic overall trend of the network behaviour, and hence any insignificant changes in the system behaviour are not taken into consideration.

In order to filter out the effects of the non-persistent state changes, the exponentially weighted moving average EWMA is used when recording the state of the system. The exponentially weighted moving average EWMA is calculated at every iteration using the formula: Ak, (-Ak -,)) +Ak ()

where: Ail ils the exponentially weighted moving average (EWMA) in cell k at time t.

Mk, is the state of the base station at cell k whose values were described above.

Î K is the smoothing constant, K =--and N is the number of samples 1+ N

in the EWMA average i. e. the number of time steps considered in the average (how many previous values of Alk are considered).

The exponentially weighted moving average EWMA provides a record with a certain amount of memory. The amount of memory the exponentially weighted moving average EWMA has depends on the value of N. For the system state at time t, elements in the matrix Fol take a 1 or 0 value dependent upon whether the average for that base station is less than or greater than a threshold value: Fkt = 1 if Akt is greater than or equal to T Fkt = 0 if Akt is less than T The value of the threshold T is set so as to determine how strict the state definition is. If T is high, then the EWMA, Ail would have to be at a high value to be recorded as a new allocation. In other words, the use of the channel concerned would then have to be very persistent for it to be taken into account in the state definition.

The process of obtaining Fk, effectively filters out the desired amount of fluctuations. A check is then performed to determine if the overall structure of the allocation pattern has changed sufficiently to be classified as a transition to a new state.

Fkt is checked against Fk (t. t) to determine how much change has happened between the two. If there is a sufficient amount of change, then Fkl is set as the new system state. If the change is insufficient, then the system state remains as Fk (t-)). This is done by calculating the percentage of elements in Fkt that is different from Fk (t-j) as follows: Gkt = Gk (t-1)if Pt is less than or equal to D Gkt = Fkt if P, is greater than D

where P, is the percentage of elements in Fk, that is different from Fett-1), D is the required percentage of change that is needed to initiate a system state transition, and Gkt is a smoothed record of the states Gk over time t.

Upon a system state transition being determined the entropy S of the system is determined using equation 1 above by considering the filtered state of each node (base station). Using equation 2, the complexity measure value is then derived from the entropy value.

Using the complexity measure to optimise backoff time in the automated setting of cell sizes The back off time (mentioned previously with respect to Figure 2) should be set as short as possible, so that the base station will be receptive to either additions or removals of neighbouring base stations. Setting the back off time too long would easily avoid the problem of two base stations continuously triggering each other off to reconfigure their respective cell size, but adds in the risk of the base station not being able to detect any changes made in the network. It is a parameter that has a significant impact on the behaviour of the network, and in a large network, it is difficult to judge its effect. We examined the effect of varying back off time using the complexity metric (measure) which has the advantage of obtaining a network-wide view of the behaviour in a straightforward manner.

When designing or optimising the network, using this metric can be done offline (i. e. non-real time) The metric is able to show the changes in the network behaviour, as a whole, which is useful when trying to discover how different network parameters and scenarios would affect the emergent global behaviour of the network.

Critical points where the network behaviour changes can be identified and avoided.

Using the metric in real time as the network is operating (i. e. online) can also be done, using measurements that are calculated over a fixed moving window to obtain the"real-time"entropy of the system.

Figure 3 shows complexity of network vs. back off time (a network parameter) as simulated using the computer model. The cell dimensioning algorithm was iterated 3000 times for each back off time value, with back off time values ranging from 5 to 37 seconds in 0.1 second increments. In polling. cell size increments of 0. 1 metres per iteration were used.

As shown in Figure 3, to the left of the hump, the network behaves randomly, to the right of the hump, the network is static after setting of cell sizes (self organising). It was observed that if the back off time as set below 14.8 seconds, the network never self-organises in terms of cell sizes and remains in a random or periodic state. Above 14.8 seconds, the network begins to gain self-organising behaviour, with the chance of reaching a self-organised state increasing as the back off time is increased.

If set above 24.9 seconds, the system would always become self-organised. This is reflected in the measured behaviour complexity, which shows the transition from one behaviour to another. The minimum backoff time which allows self-organisation is thus 24.9 seconds in this example and so 25 seconds is selected.

With the on-line complexity measurements, the network nodes are able to get some form of advance warning of unexpected behaviour. When the state space starts to increase quickly, reflected in a jump in the complexity metric values, this shows the start of a phase transition. Figure 4 shows the on-line complexity measurements displaying different behaviours when the back off time is changed to show the effects of the critical points obtained in Figure 3. The three different behaviours are observed when the back off time is set to 15,20 and 25 seconds. The network shows chaotic behaviour with a back off time of 15 seconds, while at 20 seconds, the network begins to self-organise, but isolated random behaviour starts to ripple out to the whole network, causing periodic behaviour. Finally, at 25 seconds, all the network nodes are able to organise themselves at the first time, with no undesirable interruptions. Figure 4 also shows the cell boundaries resulting from the cell dimensioning algorithm. It is important to note that Figures 4a and 4b are snapshots of the cell-boundaries, with the boundaries still pulsing, while in Figure 4c, the cell boundaries are static.

Figure 5 shows (a) on-line complexity measure values during operation of the network, back off time set to 25 seconds, one iteration being 0.6 seconds, and further base stations being added at 650 iterations. Figure 5 also shows (b) the resulting cell configuration with the further base stations added. Figure 5 (a) gives an indication of how fast the network is able to adapt to changes in the network topology. The online measurements illustrate how quickly the network goes back into its organised behaviour. The first hump in the graph shows the initial self-organisation occurring when the base stations are switched on, before reaching the desired state. At 650

iterations, base stations are added into the network (to cope with hotspots in demand, or increase coverage), triggering the existing base stations to change their configurations.

The graph clearly shows the dynamics of the reconfiguration. Figure 5 (b) shows the resulting configuration of the cell boundaries with the added base stations.

An application in respect of a wireless network has been described in detail. Of course, this development need not be restricted to use in wireless networks but can be applied in many different networks in telecommunications, from the Internet to conventional voice telephony.

The specific example described in detail relates to cell size setting. Other behaviours could be evaluated using the entropy-based complexity measure such as appropriate selection of channels at a base station so as to minimise co-channel interference with nearby base stations.

Claims

Claims:

1. A method of monitoring state of a telecommunications network comprising a plurality of nodes, each node being at a time in one of a number of node states dependent upon a predetermined set of rules and local conditions at that node, the method comprising determining the value of a parameter dependent upon entropy of the node states in the network.

2. A method according to claim 1. in which the value of the parameter dependent upon entropy is determined dependent upon sufficient node states being determined as having been changed after filtering out of fluctuations having less than a predetermined duration.

3. A method according to claim 1 or claim 2, in the parameter is dependent upon entropy S normalised by the maximum entropy possible Smax.

4. A method according to claim 3 in which the parameter is complexity

and is proportional to--1---.

? ? "ma max/

5. A method according to any preceding claim, in which the local conditions at that node are the node states of at least one of the neighbouring nodes to that node.

6. A method according to any preceding claim in which the network is a network for mobile telecommunications and the nodes are base stations.

7. A method according to claim 6, in which each base station operates according to a set of rules for automatic cell sizing defined in a local algorithm and takes at any one time one of three states in respect thereof.

8. A method according to claim 7 in which the three states are: cell sizing possible but not occurring, cell sizing occurring, and cell sizing not permitted.

9. A telecommunications network comprising a complexity determinator and plurality of nodes, each node being at a time in one of a number of node states dependent upon a predetermined set of rules and local conditions at that node. the complexity determinator being operative to determine the value of a parameter dependent upon entropy of the node states in the network.

10. A telecommunications network according to claim 9, in which the determinator comprises a filter operative to filter out fluctuations having less than a predetermined duration, the determinator being operative to determine the value of the parameter dependent upon entropy dependent upon sufficient node states being determined as having been changed after filtering out of the fluctuations having less than the predetermined duration.