WO2002087130A2

WO2002087130A2 - Adaptive distributed method for calculating system-optimal routes in packet-switching communications networks

Info

Publication number: WO2002087130A2
Application number: PCT/EP2002/004141
Authority: WO
Inventors: Paul Mathias
Original assignee: Paul Mathias
Priority date: 2001-04-25
Filing date: 2002-04-15
Publication date: 2002-10-31
Also published as: WO2002087130A3; AU2002308141A1; DE10262074A1; DE10262074B4

Abstract

The invention relates to an adaptive and distributed method for conducting a system-optimal route selection in packet-switching communications networks, which can be interpreted as a cooperation of two recurrent neuronal networks. The problem of conducting a system-optimal route selection involves the directing of data packets with a known source-destination relationship under the allocation of specified routes through the network to their destination routers as to minimize a system-wide cost function or objective function whereby adhering to certain predetermined general requirements (max. storage space of the router computer, service rates, etc.). The method is adaptive since it can continuously adapt the routes to the current load pattern of the network, and is distributed since each router computer of the network executes only local operations and, as a result, is required to communicate only with its closest directly connected neighbors.

Description

Adaptive distributed method for calculating system-optimal routes in packet-switching communication networks

Introductory remarks

Nowadays, ever greater demands are placed on the performance of packet-switching computer networks. On the one hand, the steadily increasing traffic requires more and more efficient routing procedures that optimally distribute the load over the available communication network. On the other hand, due to the heterogeneity of the services to be provided, application-related quality of service parameters must be taken into account by the communication systems. In order to respond to the various requirements of the applications and to be able to provide suitable data paths or routes, efficient routing systems are used. Routing algorithms are required to dynamically adapt the routes to the current load patterns on the network. Such routing algorithms should include various routing metrics or target functions for the application-related evaluation of routes in their calculation.

For performance reasons, so-called shortest paths are usually implemented that calculate routes through the network that are suitable for individual source-destination relationships, but without taking into account the route selection of the other source-destination relationships. As a rule, this does not lead to a system-optimal condition! The following two shortest path methods are particularly well known and have prevailed (broadly or in a modified form):

Distance-Vector Routing - The procedure is based on the Bellman-Ford algorithm. Each router periodically sends part of its routing table (destination address, distance) to its neighbors. With the help of this information, each router calculates which neighbor is currently offering the cheapest route to a destination and enters it in its routing table together with the calculated distance value. A distributed calculation of the routes takes place (no router needs to know the complete route). Overall, the process is characterized by its simplicity and low storage requirements. However, it also has serious disadvantages, such as its inertia (the so-called count-to-infinity problem), its tendency to form routing loops and oscillations, its inability to include the line bandwidths in the routing and also its slowness , Because of these disadvantages, distance-vector routing has increasingly been replaced by link-state routing. Link-State Routing - This method is considered a more flexible and robust alternative to the distance-vector routing described above. Each individual router computer has information about the entire topology of the network. The states of the system, the link states, are propagated through the entire network (flooding), so that each router knows the entire network state with a slight delay. Every router uses this data base to autonomously calculate the shortest routes for each source-destination relationship with the help of the Dijkstra algorithm. Disadvantages of the Link State Routing method are primarily the relatively high network load caused by the flooding is caused, and the high memory and processor requirements. In addition, route loops can form in certain situations (especially if the network load is included in the routing calculation). This is because the calculations are carried out independently of one another in the individual router computers.

Both algorithms calculate exactly one path for each source-target pair (single path routing).

It is therefore the object of the present invention to describe a method for calculating routes in packet-switching communication networks, which is able to determine system-optimal routes with respect to target functions which can be freely specified. The object is achieved according to the invention by a method having the features of claim 1.

The invention described here implements an adaptive and distributed method for system-optimal route selection in packet-switching communication networks with any predefinable optimization criterion (target function). For this purpose, the actual communication network and the associated routing method are mapped to special, mutually cooperating neural networks, which are called the identification and error propagation network.

FIG. 1 shows that the task can be interpreted as a control problem in that the actual communication network is understood as the system to be controlled (identification network) and the optimal route selection as control of the first subsystem. Structures and states of the identification network are defined by the topology of the communication network and its specific attributes (queue memory sizes, service rates, etc.). The free parameters of the identification network to be set are the entries in the routing tables of the individual router computers. The controller (the error propagation network) sets these parameters according to a given target function. Both subsystems are constantly cooperating with one another: the identification network cyclically transmits its current system status, while the error propagation network cyclically returns the new corrections to the routing tables.

The optimal routes are determined with the aid of the error propagation network, which as a linear recurrent neural network implements a modified version of the particularly efficient learning method recurrent back propagation. The calculations of the error propagation network are all local, i.e. that each router computer only has to carry out a small part of the total calculation. The model is capable of routing any subsets of data traffic (e.g. traffic to a destination) separately in so-called layers, taking into account the interdependencies of the layers.

The method is suitable both for connectionless intermediary layers (ie data program subnetworks such as TCP / IP) and for connection-oriented services (subnetworks with virtual connections such as ATM), or for mixed forms of the two. It is simple, robust and stable. Additional criteria such as fairness and optimality can be introduced via suitably defined target functions. Solution of routing problems with the help of recurrent neural networks

Mapping of the topology and dynamics of a routing network onto a recurrent neural network

A routing network is defined by a directed graph $ = fe, S), with "K - set of N router computers and S = set of M router connections," router connection "here for the transmission channel together with the corresponding router output and its queue memory (output queue)! Q and Z are the number of data sources or data targets in the network. Be further. c 5 the set of source connections and cz S the set of target connections.

First, a first recurrent neural network (RANN) is constructed, which realistically reproduces the topology and dynamics of traffic in the communication network. This network should, according to the usual terminology of control models, identification network or

Identification RANN are called. The topologies of the communication network and the neural network are mapped to one another by the router connections S of the communication network with the

RANN neurons can be identified. The connections of router outputs realized by the communication network and its router computers thus correspond to the couplings of the

Neurons. The network thus defined is a recurrent neural network, since in principle every one

Neuron can be connected to each other. In this case, however, the neurons are sparsely coupled, since a router is usually only connected to a few other routers.

Note that there can be no two-way connections between two neurons because the graph G lines are directed.

Figures 2 and 3 illustrate the described assignments of the topologies using a simple example. 2 shows a simple communications network is shown with router computers (__ to R _4), data sources _(Q;, Q _2), data targets _(Z, Z _2), and connection lines with the associated queue memories (__ _/ to L, ₂ ). For example, the router R _{3 has} the option of forwarding the data packets arriving via line L ₅ via lines L ₆ , L, ₀ or L ,,. The figure shows the mapping of the communication network from FIG. 2 to a recurrent neural network. The individual neurons (N _/ to N ₁₂ ) correspond to the connecting lines with the associated queue memories of the communication network. The entries in the routing tables of the router computers which define the multipath routing of the communication network are the coupling weights w _{Lk of} the neural network. The coupling weights must add up to 1 (e.g. w + w, = 1).

Due to the described identification of router connections and neurons, the activity state x, (t) of the .th neuron at time t corresponds to the number of data packets that are in the queue of the router output ie S at this time. The propagation of the data packets in the communication network is the neural network with the aid of specific activity functions s "r _t and coupling weights w _lk, i, k S is realized. These functions, which depend on the state of the neuron (or the number of data packets), are interpreted as service (_,) or as memory restrictions (r,) of the communication system. You determine how the neural states change per time step, or in the language of the communication system, how many data packets reach or leave a router output memory in a time step. The coupling weights w _ik determine how the data packets arriving at the router inputs i are distributed to the router outputs k. You use it to define a routing table. To the For example, a weight w _lk = 0.5 means that the router containing the router output k routes exactly half of the data packets arriving at its router input from the router output of another router computer to the router output k. For reasons of data conservation, the weights of the RANN are ≤ w _tJ ≤ \, V, y and, ^w „= ^l > ^v (0.1)

required.

However, the identification RANN constructed in this way is too unspecific as a partial model for a routing method, since it cannot yet take into account the source-target relationships of the data packets. These relationships are formally given by time-dependent data traffic matrices U (t) = u _lk (t), dim (U) = QZ. The matrix entries u _lk (t) indicate how many data packets per unit of time have to be transported from the source i ≡ O to the destination k & $.

The inclusion of the source-destination traffic requests in the model is done by expanding the identification RANN. For this purpose, the identification RANN is made up of Z individual layers, each of which is a copy of the original network topology. Each layer is supposed to simulate the partial flows of the overall scenario, which, starting from any data source, lead to exactly one goal. The number of layers of an identification RANN expanded in this way is identical to the number Z of the destinations of the communication network. In this way, a specific subset of the total data traffic is modeled in each layer. In order to avoid dead ends, all neurons and connections that are not on any route leading to the destination d (from where the destination in question cannot be reached) are still removed in each layer d. FIG. 4 (a and b) illustrates how an identical copy of the original neural network is constructed for the sub-data streams of each target. These copies are called layers. With the introduction of the layers, all state variables of the original network are also multiplied. The new sizes are distinguished by a top index (e.g. w ₃ ² ₇ is a coupling weight of layer 2). The figure shows the two layers for the example from FIGS. 2 and 3. In layer i, exactly the data streams of the overall system that have the target i are then modeled. However, it should be noted that the layers are not completely independent of one another, since the dynamic equations (0.2) of the neural network also include sums of layer-specific state variables across all layers. In each layer, the dead ends to the other targets are removed (ie the coupling weights of these connections are constantly set to the value 0).

The source data flows q _k ^d of each layer d can easily be calculated either from the specified data traffic matrix (q _k ^ά - = u _ω ), or read directly from the statistics of the relevant source router.

The following applies to the extended identification RANN: 1.

Output; ^' with the aim d. For the rest, let x _l : = (x], x ^, ..., xf) be the vector Data sets of all layers, x ^rf : = (x ^d , x ^d , -, x _M ^d ) the vector of the data sets of layer d, and x: = (x ', x ² , ..., x' ^! ) The vector all amounts of data.

2. The data flows of the individual layers are not independent of each other. As will be shown later, both the service s ^d and the storage restriction r _t depend on the activity states x of all layers.

The discrete-time dynamic equations of the individual layers of the identification RANN are given by x ^d t + τ _x ) = x ^d t) + u _k ^d {t) -hi {t) + τ _x q _k ^d (t)

- ^ (t) + r _t (X v _I l (t '' (t) - ^ ( ^/ ) £ ^ (t) r, (t) + τ _A ^rf (t) - ⁽⁰ - ²⁾

The RANN defined by equations (0.2) is a so-called Σ-II network because it contains weighted sums of products of (sigmoid) functions. Here and below, the top index d always denotes the number of the shift. The q are the source data flows of layer d, and τ _x is the discrete time interval of the development.

The service s ^d of layer d is the part of the total service s, which corresponds to the ratio of the amount of data x ^d of layer d to the total amount of data x _t :

^ (x,): = s {x) ^, with s ^d (x,): = 0, if *, = 0 _(0. ₃ )

The total service s, (x _t ) e [0; _,] describes the total transmission power of the router computer via the connection i depending on the amount of data x, the associated output memory, where c, is the maximum service rate for the output memory /.

The form of such a function is shown in FIG. 6. Here is the rough functional form of the total services s ^ x,) plotted over the number of packets x, the queue memory. The transmission power s is limited by the service rate c. The exact definition of _ is determined in a specific application by a number of other (fixed) parameters, such as the performance of the computer hardware and the channel capacity of the line. The method described here can use any service functions s _t as long as they are monotonously increasing and continuously differentiable. An example of s, is _, (x,): = 2 τ _x c, (l + exp (-α, x _t )) - 1 with free parameter α,. The monotonically falling r _k e [0; 1] are the memory restrictions. They express the inability of a full output memory to accept further data packets. The r _k depend on the total amount of data in the output memory k, ie r _k (t) = r _k (x _k (t)). Requirements for the r _k are: falling monotonically, differentiable and a range of values between 0 and 1. An example for r _k is.; (.;,): = L- (l + exp (α, -b _l (x _l lx _maxj ))) ^'1 with free parameters a _t , b, and a data storage size x ^ _,. The exact definitions of the r _k are insignificant for the method as long as they meet the above requirements.

The form of such a function is shown in FIG. 5, the memory restrictions.; (;.,) Above the Filling level xx _max ,, of the reservoir are plotted. Up to a certain fill level ( ^xx _maι ., - ^< l), all incoming packets can be picked up, afterwards an increasing number of packets have to be discarded until finally, with full memory (x, / x _maxι = 1), no packets can be picked up. The definition of the dynamic equations (0.2) shows that the real incoming and outgoing data streams u ^d (t) and h _k (t) depend on both the service rates and the memory restrictions of the router computers involved. In the case of the outflowing data stream h _k (r), this does not quite correspond to the real situation, since packets that cannot be received by the subsequent routers and their output memories must be lost and sent again, whereas this simulation model assumes that data packets that cannot be received by the following router computers are not sent at all. It is crucial in this context, however, that the later appropriate error propagation network calculates backflow-free routes that automatically minimize the loss rate and the resending of data packets. With the introduction of the correlation terms a ^d _k (t) = s ^d (t) r _k (t), (0.2) can also be written as follows: ^x i (t + τ _x ) = x ^d t) + ∑ (t) a ^d _k (t) - _j w _h ^d (t) aii: t) + τ _x q _k ^d (t). _{(Q 4)}

The a ^d _k will be used later for further calculations.

The construction of a cooperating error propagation network

The error propagation network is the routing model part of the process. In it, error signals are propagated in the opposite direction to the data packets of the identification network. The states of this linear recurrent neural network are used to calculate corrections Aw _{k of} the weights w ^d _{k of} the identification network according to predetermined optimization criteria (the w ^d _k determine the routes of the data packets through the communication network).

The error propagation network implements a modified version of the learning method recurrent back propagation via a gradient descent along a specified target function. The development of the required formulas is shown here in a shortened manner - without intermediate steps - since the complete explanations would go beyond the scope of this description of the invention. In principle, however, the explanations follow the well-known derivation of the recurrent back-propagation lem method, as it is e.g. is explained in [4]. The main difference is that the identification network is a so-called Σ-11 network.

Depending on the application and the desired optimization goal, the target functions can be defined almost arbitrarily (they must be differentiable in x _t ). Useful target functions for optimal routing procedures in communication networks are, for example:

E (= ∑x, (s, (xXt)) ,, (*, (,)) = ∑ _{C |} . ^)) • (0-5)

with an M / M 1 waiting time function π, (s) = s / (c, -s), or simply: E (t) = φ, xt), (0.6)

with the positive weighting constant φ _t , or a function that forces the same transport times for all alternative routes of a source-destination pair:

with an arbitrary delay time function π, (s).

For the correction of the weights follows the gradient descent method

with a step size ε of the descent and the constants derived from the objective function used! ^ = dE _k / dx _k . For example, for the target function (0.5):

_, 2c _k -s _k * = * _* * _* rr- (0.9)

The unknown partial derivatives dxl / dw ^d are determined using the dynamic equation (0.4); it results

With

a <

3

(lim _ ^ (._,) = 0 and limG, _Jl (Λ :,) = l / π (0)> 0). Conversions and substitutions of (0.10) result in a new linear RANN (the error propagation network) with the dynamic equation

and coupling weights

However, the linear system defined in (0.12) is not contracting, since the matrix v ' _k "has a spectral radius greater than 1. A conversion of the equations into the form

∑ ^v 7 _k y = -∑∑ y, '(0.14) ι "t / l ≠ e, however, allows the system to be interpreted as two cooperating error propagation networks. The left-hand side of the equation defines the 1st error propagation network with a weight matrix v ". Its task is to solve the linear system of equations (0.14) iteratively with the right side fixed. The second error propagation network with the weight matrix v ^ _k, on the other hand, solves the system of equations (0.14) with the left side fixed. So the dynamic equation of the 1st linear network is

(0.15)

while the 2nd network realizes the following linear operation

; (' ⁺ *,) = Σ∑ - ∑, (') - + ∑∑ * Jtf ((0.16)

The two linear networks work alternately until a stable state occurs (the solution y ^{* of} the system of equations defined by (0.12)). With the help of the solution y ^* , the required corrections of the weights can be determined as follows:

Δ = ^β (- tf) _β (0.17)

Mode of operation of the overall model The mode of operation of the overall model consists of developing the identification network and the two error _{propagation networks} alternately or in parallel in accordance with their dynamic equations, and after an adjustable number of time _cycles t _ldenl or t _error, the current network states are sent to the other RANN to calculate its weight corrections to hand over. The integer timing _parameters t _ldent and t _error determine how often corrections are made to the weights and how accurate these corrections are. They can be integrated as a kind of inertia of the system (larger values correspond to a larger inertia).

The following steps are carried out cyclically either in succession or in parallel: 1. Identification network: _Develop the system state over t _ldeπl discrete time steps of length τ. according to the dynamic equations (0.2),

2. _Error propagation network: Take the current state x ^d (t) of the identification network and calculate the weights v according to (0.13), 3. _Error propagation network: Develop the system state via t _error discrete time steps of length X _y according to the dynamic equations (0.15) and ( 0:16)

4. Identification network: Take the current state y (t) of the error propagation network and update the weights w ^d _k using (0.17).

Figure 7 shows an overview of the overall model. It can be seen that the overall model is essentially divided into two: the layered identification network with the functions s, (total services) and r ₍ (memory restrictions), both of which depend on sums of state values of the individual layers , and the two-part error propagation network, which realizes the system-optimal routing. The error propagation subnets receive weights v | or v * calculated cyclically from state variables of the identification network and return weight corrections for the coupling weights of the identification network Routing tables of the communication network are dynamically adapted to the changing conditions (network load and traffic refueling).

Transfer of the RANN routing model in the context of packet-switching communication networks Implementation of the RANN routing model in individual router computers

Two determinations are decisive for the implementation of the RANN routing model as a procedure in individual router computers:

1. Both the layered identification network and the error propagation network can be mapped directly onto the communication network, since all have the same topology. For this purpose, the individual router computers are assigned the neurons or layer neurons that correspond to their output memories. The services s ^d and storage restrictions r _{t of} the identification network, which are used for the calculations of the fault propagation network, can either be defined in such a way that they correspond to the performance characteristics of router computers and connecting lines, or the required function values and their derivations are simply provided by the router - Calculators kept as statistics and used if necessary.

2. All occurring arithmetic operations of the fault propagation network are local, i.e. they can be distributed to the individual router computers in such a way that each router computer only has to carry out a small part of the total calculation and only needs its own status data and that of the directly connected neighboring router computers. This creates a completely distributed process.

The concrete sequence of the procedure in a packet-switching communication network is divided into four function blocks, which work in parallel, and each block has its own fixed execution cycle: 1. Route data packets and create statistics

Each router computer routes the incoming data packets according to its routing table, which is formed from the weights w ^d _k . The weights with the index d are responsible for the packets with the destination router d (or for a correspondingly defined partial stream d). The multipath routing defined by the w ^d _k (usually exist for fixed i indices j "j ₂ with ^w _j , ' ^w _jι ^> 0 wi ¹ ^ realized by the data packets in a corresponding ratio to the outputs _ /, and j _2. This division should largely take place in such a way that the data packets of a source IP address are assigned to exactly one output. Furthermore, each router maintains k statistics (mean values over time) about all its partial data streams s _k , s _k , their temporal Changes s _k 'and the lengths of its output queues x _k , x _k .

2. Propagate status variable and calculate v-weights

The statistical values of the router are sent cyclically to all directly connected neighboring router computers after defined time intervals τ _Statm . After receiving this data, each router is able to calculate its error propagation weights v *, v * according to the formula (0.13).

3. Calculate and propagate error signals

After fixed cycles τ _Errl and τ _Err2 , each router calculates its error signals y _k and z ^d according to the dynamic laws (0.15) and (0.16). The new error signals are sent to all directly connected neighboring router computers. 4. Calculate and propagate weight corrections

After a fixed predetermined time pulse τ _Kor , each router k calculates the corrections Aw, Aw _{k of} its routing weights w _h ^d , w ^d _k using the current values y ^d according to the formula (0.17), and normalizes the new weights. The new weights are sent to all directly connected neighboring router computers.

Advantages and special performance features of the process

1. The model is a system-optimal routing method, optimal with regard to arbitrarily definable target functions. System-optimal also means that the method does not calculate the individual routes independently of one another; in contrast to link-state routing, where each router computer calculates its shortest paths independently of the calculations of the other router computers (which can cause oscillations). The routes are generally calculated depending on the load, the route patterns permanently adapting to the current traffic demand and the status of the output memory.

2. The calculation method for the dynamic determination of the optimal routes is particularly efficient since the back propagation algorithm is currently the fastest way to calculate gradients for a gradient descent method.

3. The model is a distributed process in which each individual router in the network only carries out a small part of the total calculation. The steps in the faulty propagation network break down into elementary operations, all of which are only local information - one's own and that of the next neighbors - need. The local operations are simple and do not place any special demands on the individual routers in terms of computing time and memory requirements (in contrast to link-state routing). The additional communication effort for the error signals is low. The individual router computers do not need to know the entire topology of the communication network.

4. Although the underlying theory with the recurrent neural networks appears to be relatively complicated in terms of mathematics, the local computing operations for the individual router computers are particularly simple in the concrete implementation.

5. The model implements multipath routing (w ^d _k ^ [; l, _k w _k = 1). It routes the information packets of a source-destination relationship in the event of high network loads, also via different, alternative routes, so that the available network capacities can be used to the maximum. The data packets can be divided into different routes, for example, source-specifically. A multipath routing can neither be realized by the distance vector routing, nor by the link state routing. 6. The model is robust against router failures. A failed router computer can no longer participate in processing the error signals. As a result, the turn rates of the neighboring computers that affect him are reset immediately (or very quickly). This is tantamount to automatically reducing the topology around this failed router. The opposite process is also possible: a router computer is newly integrated into the network, establishes a connection with its neighbors and is thus automatically included in the global routing strategy without the new computer having to be announced to all computers in the communication network.

7. The model is able to provide heterogeneous services (prioritization, guarantee of quality-of-service, multicast routing or mixed single and multipath routing), because on the one hand, any part of the total traffic is modeled and optimized in the layers of the identification network and, on the other hand, each individual layer can be provided with different dynamic parameters and target functions. The support of such specific services cannot be achieved with the usual methods (distance vector routing, link state routing).

Publications cited [1] P. Mathias, Static and dynamic traffic allocation with recurrent neural networks, Shaker Verlag, ISBN 3-8265-6720-X, 1999.

[2] M. A. Sportack, IP Routing Fundamentals, Cisco Press, ISBN 1-57870-071-X, 1999.

[3] AS Tanenbaum, Computer Networks, Chapter 5, Prentice Hall, 3rd edition, ISBN 3-8272-9568-8, 1998. [4] J. Hertz, A. Krogh, RG Palmer, Introduction to the Theory of Neural Computation, Chapter 7, Addison Wesley, ISBN 0-201-51560-1, 1991.

Claims

claims

1. Method for calculating, with regard to predetermined network-wide target functions, optimal routes for the data units to be transported from a source to a destination in communication networks, by means of a data processing system, characterized by the following steps:

• The topology and dynamics of the communication network are mapped onto a special recurrent neural Sigma-Pi network (identification network) by assigning output memories of routers arranged in the network together with subsequent connecting lines to the neurons of the neural network, which results in activity states (x _k ^d ) of the neurons the number of data units of a target (d), which in the corresponding output memories on their

Waiting forward, correspond, and coupling weights (w ^d _k ) of the neurons are the portions of the incoming data units with destination (d) that are routed to the possible outputs (routing table), the Sigma-Pi network being constructed from individual layers , in which partial data flows are mapped, »the router computers determine data throughputs (s _k , s _k ), memory restrictions (r _k ) and queues in the output memories (x ^d , x _k ) as well as changes over time (x _k ', s _k ' , r _k ) thereof and transmit these variables at time intervals (τ _stato ) to directly connected neighboring router computers,

• For the purpose-dependent routing, the router computers calculate a fault propagation network on the basis of a version of the learning process “recurrent” adapted to the Sigma-Pi network

Backpropagation ", whereby each computer only carries out the part of the overall calculation that relates to its local status variables;

- each router calculates at intervals (τ _Λαήβ ) the weights (v *, v *) necessary for the calculations of the faulty propagation network, - each router calculates its error signals (y _k ) and at further intervals (τ _{£ rrl} and τ _Err2 )

(z ^d ), after which the new error signals (y _k ) are sent to directly connected neighboring router computers,

- After an additional time interval (τ _Kor ), each router calculates corrections to its routing weights (Δw, Δv) using the current values of the error signals (y _k ^d ) and normalizes the new weights, after which the new weights are sent to the directly connected neighbor. router

Send computers that use these new weights as a new routing table.

2. The method according to claim 1, characterized in that the communication networks are packet-switched.

3. The method according to any one of claims 1 or 2, characterized in that that the occurring time intervals (t _Slalus , τ _Erτl , τ _Err2 , τ _Kor ) are _tracked alternatively, for example depending on special events or the load of the communication network, so that, for example, the additional communication effort is minimized.

4. The method according to any one of claims 1 to 3, characterized in that certain connections or routes of the network are more attractive than others due to additional weight factors in the target functions used and are therefore preferred for route selection.

5. The method according to any one of claims 1 to 4, characterized in that the partial data streams of a different target are assigned to each layer.

6. The method according to any one of claims 1 to 5, characterized in that to include heterogeneous prioritization services for connection-oriented services, the individual layers are assigned freely definable partial data streams (e.g. a single source-target relationship) which are to be treated and routed separately.

7. The method according to any one of claims 1 to 6, characterized in that specially marked data units of a source for claiming a particularly prioritized route for connection-oriented services on the shortest route are routed through the communication network and for himself and the subsequent data units this path by new formation Reserve high priority shift.

8. The method according to any one of claims 1 to 7, characterized in that a certain subset of the routes in the communication network is fixed by constructing layers, the weights of which are fixed, ie are not changed by the error propagation network.

9. The method according to any one of claims 1 to 8, characterized in that the router computer, the data throughputs (s, s _k ), queues (x _k , x _k ), memory restrictions (r _k ) and the temporal changes thereof (x _k ', s _k ', r _k ) determine over time averages and make them available for further calculations of the method.

10. The method according to any one of claims 1 to 9, characterized in that the router computer, the data throughputs (s _k , s _k ), memory restrictions (r _k ) and the temporal changes thereof (s _k ', r _k ) via predefined functions determine and make available to the further calculations of the method.

11. The method according to any one of claims 1 to 10, characterized in that the router computers in the calculations of the Fehleφropagierungsnetz simplify all coupling weights that relate to different layers to 0 (vj; '= 0, e ≠ /), so neglect ,