Embodiment
In Fig. 1, the flow chart of optimizing the automated procedure of network performance according to first embodiment of the invention is shown.At first, at step S1, select to be used to describe the key performance indicators of performance of the part interested of network.Then, at step S2, determine the configuration parameter of KPI based on it.At step S3, selection will be included in the quantity of the sub-district in the optimizing process,, selects the sub-district cluster that is.At step S4,, determine the currency of KPI according to each configuration parameter.After this, at step S5, determine the currency of KPI according to the quantity of the currency of KPI and sub-district, function assesses the cost.At step S6, the result of cost function, value and the configuration parameter of KPI are stored in the historical data base.
At step S7, at least a value in each configuration parameter is adjusted, obtain one group of new configuration parameter.Organize new configuration parameter according to this, determine new KPI value, then,,, recomputate the cost function according to new KPI value and the number of cells (changing) selected at step S3 at step S5 at step S4.At step S6, also the new result of cost function, new KPI and configuration parameter value are stored in the temporary storage.Subsequently, at step S8, will be based on newly/this new result of the cost function of the configuration parameter group adjusted compares with the previous result who is stored in the cost function in the historical data base, to determine after the adjustment configuration parameter, whether interested network performance improves.
If after adjusting configuration parameter, network performance improves,, the configuration parameter group of adjusting is used as permanent parameter then at step S9.Yet if after step S8 determines to adjust configuration parameter, network performance does not improve, and at step S9, will be used as permanent parameter in first set of configuration parameters that step S6 stores in the historical data base.
At step S10, whether the at interval interior at the fixed time network performance of check improves.When at the fixed time at interval in network performance when not improving, that is, even automatically adjust, KPI history does not improve yet, goes wrong at the automated procedure of optimization network performance step S12 informing network operator.Owing to obviously can not regulate many parameter values automatically, and adjusting can not be optimized network all the time automatically, so whether the operator can check this problem to be because hardware problem, and perhaps under the current network condition, whether cannot the Automatic Optimal network performance.Be in this case at network, network operator must manually be optimized network performance.
On the other hand, when at the fixed time at interval in network performance when improving, flow process jumps to step S7, at step S7, adjusts configuration parameter again, so that further optimize network performance.Then, this flow process will continue flow process described above.
In a second embodiment, not only select relevant KPI, and determine one group of QoS target, utilize one group of benchmark KPI to represent this group QoS at step S1.The automated procedure of optimizing network performance according to second embodiment and optimizing process basically identical according to first embodiment.Unique difference is when when step S5 assesses the cost function, to utilize the difference of KPI and benchmark KPI to replace the KPI value.
Therefore, the operator is provided with capacity requirement to the specified vol KPI that utilizes " ref " in the subindex to be expressed as KPI_C.Correspondingly, the operator is provided with quality requirement to specific KPI_Q.Then, utilize equation (1) calculated mass cost and capacity cost.
Utilize different weights factor alpha and β, with different cost function combinations or added together.By control or change weight coefficient α and β, can emphasize the cost and the general status of particular type.
Can think that its task is the combination that the mathematical formulae that is used to optimize network performance is used for according to which KPI determining near the zone that requires as far as possible the air interface configuration parameter.
Fig. 2 illustrates the example of KPI cost function f.In this example, be higher than linear the increasing of KPI value at cost of KPI_ref.
Equation (3) illustrates and will optimize, and promptly will be reduced to minimum total cost function.Utilize parameter W, can between capacity and quality requirement, realize balance.By adjusting configuration parameter (2), can be reduced to minimum.The KPI value also depends on service assignment, for example, obtains different cost setting and parameter setting according to service assignment.
KPI_C
i=f (configuration parameter, service distribution)
KPI_Q
j=f (configuration parameter, service distribution) (2)
The * capacity cost (3) of total cost=W* quality cost+(1-W)
For example, the factor that can influence optimizing process is: the price of service distribution (service mixes), traffic density, every kind of service etc.The final goal that total cost is reduced to hour comprises: Optimizing operation person's income, CAPEX and OPEX be reduced to minimum and attended operation person's favorable image.
According to following equation (4) to (8), the specific examples of the function T that can assess the cost OTAL COST:
(4) total cost=C (queuing ratio)+C (poor quality ratio)+C (call drop ratio)+C (blockage ratio)
Wherein
(5) C (queuing ratio)=0.05*Dev (the queuing ratio that queuing compare-is allowed)
(6) C (poor quality ratio)=0.2*Dev (the poor quality ratio that poor quality compare-is allowed)
(7) C (call drop ratio)=1*Dev (the call drop ratio that call drop compare-is allowed)
(8) C (blockage ratio)=0.10*Dev (blockage ratio-allow blockage ratio)
The optimization task is exactly that the TOM management level that all are different are combined, and should consider that wherein the measured value (quality status stamp and cost mark) of different layers uses this fact of different language.
When carrying out optimizing process in the NMS at network, support the operator to judge customer service and each SML.In order to realize this process, will comprise configuration and the deviation of the cost function measured at (PM) that lower floor carries out is converted to " language " on upper strata.This can realize by carrying out following content:
Execution is converted to (reflection) quality expection/target relevant with service from radio access network parameter (setting).In fact, this means, make configuration management relevant with performance management.That is, utilize one group of particular measurement value, the functional entity that monitoring has customized configuration.Utilize the cost function that adopts the measured value of determining to calculate the performance of this entity.
In fact, mean below and say that to delivering to the measured value (that is, sub-district, class of service etc.) the user class entity, big entity of the quality that can determine each user changes from statistics.Also utilize (respectively) weighted cost function, carry out this step.In addition, these conversions can be combined with cost function, with the end user's quality indication that realizes requiring.
1) the technology conversion (reflection) of end user's flow process rank (experiencing quality) from radio access network measured value (network performance) to service.
2) from assembling the technology conversion (reflection) that level (UMTS class of service) parameter is set to end user's flow process rank (experiencing quality) of service.
3) the technology conversion (reflection) from the measured value of each class of service of being provided with to end user's flow process rank (experiencing quality) of service.
And/or the composite function of class of service and process level information (parameter and setting) and cost function, with support parameterization and monitoring end user's GOS.
Fig. 5 illustrates the schematic diagram of the combination that is used to utilize network monitor function that reflection combines the different management level in the network and majorized function.
By network measure value, performance indicia PI and/or KPI and cost function are combined, can realize reflection from one deck to following one deck.
Can calculate the cost function of the grade of service GOS of user's impression, as equation (9) expression:
(9) GOS=C (service availability)+C (postponing and vibration)+C (quality)+C (call drop)+C (service accessibility)+C (equivalent bit rate or user throughput)
Postpone wherein to comprise that service access postpones and the queuing transmission delay.The non real-time quality is subjected to the influence of packets lost, radio link control RLC, packet data convergence protocol PDCP,, is subjected to the influence of error rate BER and BLER (block error rate) BLER that is.About real-time quality, if up link UL BLER (block error rate) BLER is significantly higher than target BLER, then this quality bad luck.Quality is subjected to down link DL to connect the influence that has a power failure in real time.The income of above-mentioned cost function comprises capacity requirement and service distribution.With kbp/ sub-district/MHz is that total throughout is measured by unit.
When the user 98% was satisfied, it was the throughput of unit that the spectrum efficiency of cost function equals with kbp/ sub-district/MHz.This means that professional accessibility and blocking probability are 2%.The equivalence bit rate is higher than 10% of load service data rate, and 98% user is not by call drop.Adopt the reason of this method to be, on tolerance, help being optimized according to GOS.
Must promptly, utilize the service of different parameters setting or other property control to carry out this reflection to all services that provided.
Although each conversion all makes precision reduce, from statistics, reflection is correct.Because the statistics level can be operated, so the optimum position of reflection function is at NMS.In addition, the NMS implementation procedure can also be handled radio network controller RNC-RNC (or other network element) borderline region.When carrying out this conversion, use the cost function of suggestion at every turn.Therefore in some cases, the service QoS target can cause the parameter setting to clash, and needs the cost function problem that manages conflict.This can realize by in cost function different units being provided with different weight coefficients.When entering network system, this theory is extremely important in different client's classification (silver, copper, gold etc.).
In addition, when being altered to last management level of TOM model, next key step is to calculate to the network optimization, SO service order and with the client area partite transport that Euro, dollar or pound are represented to calculate.In this stage, need interior bill and the pay imformation that inventory/the gathering subsystem provides of customer service layer of TOM.When the knowledge of the characteristic of utilizing client's basic condition/distribution and these distributions, can operator's commercial situation be optimized to the most useful direction according to cost function.It should be noted that to change client's priority and, will cause client's behavior to change, and therefore repeat Business Management Layer's optimization because of the QoS that commercial reason provides.
Have optimum performance in order to ensure cellular network, the operator preferentially makes flexible apparatus according to KPI of system (key performance indicators) and/or thus obtained cost function the QoS target is set.Can perhaps the QoS target be set to the sub-district cluster to each sub-district.Can according to because the blocked call of hardware resource, " soft " blocked call (interference-limited network), call drop are called out, quality is badly called out, for the repeating transmission number of times of packed data and delay, diversity switching probability, direct-cut operation success rate, load condition (up link UL or down link DL), to the packed data of circuit switched service etc., can calculate QoS.
Under the multi radio environment (GSM-WCDMA global system for mobile communications-Wideband Code Division Multiple Access (WCDMA)), importantly, for optimizing capacity, convergence and quality, can be to these two network creation resource pools.This requires that higher (KPI) layer is had all controlled function (quality manager), that is, utilize quality manager can realize according to optimizing process of the present invention.
Quality manager QM, that is, optimizing process provides the center monitors function, and the state of monitoring parameter value, compares the automatic recognition problem situation by the historical information that will be stored in the parameter value in the historical data base.For example, can be as far as possible little, as far as possible independently GERAN and UMTS terrestrial radio Access Network UTRAN are divided into automatic adjusting subsystem.By KPI weight coefficient is set, considers interdepending between the subsystem on the upper strata of quality manager to its each subsystem.
In another embodiment, according to user's group (as business users, free time use etc.), carry out optimizing process.
Can put it briefly the default value of current all parameter values of suggestion.Up to the present the community user network is still operator's task (attempting to consider many cell environments) one by one.Yet, adopt the method and/or the system that optimize network performance according to the present invention, it is inessential to make initial parameter be provided with.For example, when network brings into operation, can under the restriction of very " pine ", carry out access control and switching controls, thereby according to current QoS situation (being positioned at the KPI of the OSS of operate services system) with can regulate QoS target related parameter, that be provided with automatically, make all user access networks.In the parameter change new situation, after the promptly new KPI value, the historical data of it and KPI is compared, and if the change of QoS performance (the perhaps cost function of qos requirement) improve acceptance " test " parameter then.The length of historical data depends on the traffic carrying capacity (the sampling sum should be enough high) on the network.Importantly, the QoS cost function contains the projects in whole RRM and multi radio zone.
Currently at first key parameter (according to optimum capacity and quality) is set to " default " value, in most of the cases, should " default " value guarantees the network operation, but be optimum performance.Optimizing process according to the present invention changes to best operating point according to overall QoS automatically with being provided with of basic parameter.
The regulated quantity of configuration parameter can fixed increment or decrement.As a kind of selection, increment or decrement can be variablees.
A third embodiment in accordance with the invention is utilized cost function to concentrate and is optimized Internet resources, and the service quality (QoS) of calling hierarchy is provided.Cost C is the function of the variant KPI of network, for example:
KPI wherein
iBe i key performance indicators, and Fi is certain positive function that can be used for conversion, weighting and/or i KPI of convergent-divergent.By this cost function C is reduced to minimum, optimize network performance.Be counted as the variant network parameter W=(w of the optimum value of parameter by correct selection
1w
2..., w
N), can realize least cost.The cost function method supposes that impliedly the value of KPI is the function of network parameter, that is:
KPI
j=KPI
i(w
1;w
2;...,w
N)j (12)
Therefore, cost function C still is the function of network parameter, and it can be rewritten as the direct function of parameter:
G wherein
iIt is certain function of t variation in time.
The 3rd embodiment relates in particular to a kind of simple, efficient algorithm that is used for above-mentioned cost function is reduced to minimum.Yet, under actual conditions, optimize this cost function and do not lead directly to.List its subject matter below:
1. in real network, there are many kinds of types of service, user distribution and loads.Network can not be controlled these factors, and these factors can be regarded as external noise source at random.Any optimized Algorithm all should be inresponsive to this random external influence.
2. because the differences such as load of different time are selected (that is function g of equation (13), so can improve the optimal parameter of any given network at any time
1Change in time).Any optimized Algorithm all should adapt to this variation of network best operating point, and can follow the tracks of these variations.
3. may not produce the network model that can under various different situations, be optimized.This means, do not know the function gi of equation (13).The replacement method that the 3rd embodiment adopts and this method be any model of hypothetical network not, but makes optimizing process only according to the network measure value.
The optimized Algorithm of using in cost function being reduced to minimum process is described, and can realizes this optimized Algorithm according to first principle.Study cost function C being reduced to minimum ordinary circumstance according to the parameter that is represented as w.If w
0It is the value that is used for C is reduced to minimum w.Utilization is calculated C (w about the taylor series expansion of any value of w
0) draw,
Wherein C ' is the single order differential of C to W (w), and C " (w) be second-order differential.Because C (w
0) be the smallest point of C, about w
0, ask the differential of equation (14), and establish the result and equal 0, then draw,
It is classical Gauss-newton (Gauss-Newton) algorithm optimization algorithm of restraining fast.If C is the quadratic function of w, then a step converges to optimum w
0If C is not a quadratic function, then as long as C " (w) be positive all the time, just can guarantee to restrain.If it is not, then can make it for just, and equation (15) is decomposed into the normal gradients algorithm.Yet owing to the smallest point near C, so second approximation is more accurate, and convergence rate is faster.
Now, according to the WCDMA cost function, study this problem.As mentioned above owing to there is not the model of network, so be difficult to determine C ' (w) and C " (w) value, therefore be difficult to use equation (15).Yet, according to the 3rd embodiment, utilize the network measure value calculate C ' (w) and C " (w) value so that use equation (15).
Minor variations δ w>0 to the value of parameter w is studied, to draw new parameter value w+ δ w, so the value of cost function can be approximately
Equally, can be in the hope of expression formula C (w-δ w):
These two expression formulas are added together, and rearrange everyly, obtain C " (wadd), deduct these two and rearrange this expression formula then, produce C " (wadd) expression formula.
Therefore, by knowing C (w+ δ w) and C (w-δ w), can utilize equation (18) and (19) try to achieve C ' (w) and C " (w) value, perhaps their approximation.Next problem is how to calculate these values at any special time.This is to carry out according to following step:
1. at time t1, parameter value is w, and according to the network measure value of the correct KPI of time t1, utilizes equation (11) function C (w that assesses the cost; T1).
2. at time t1, the w value is changed to w+ δ w.
3. at time t2=t1+ δ t; δ t>0 according to the network measure value of correct KPI, utilizes the assess the cost value of function of equation (11), to draw C (w+ δ w; T2).
4. at time t2, parameter w is changed to w-δ w.
5. at time t3=t2+ δ t,, utilize equation (11) function that assesses the cost, to draw C (w-δ w according to the network measure value of correct KPI; T3).
6. at time t3, utilize respectively equation (18), (19) draw C ' (w) and C " (w); and utilize measured value C (t1), C (t2), C (t3) to draw under the situation of C (w), C (w+ δ w) and C (w-δ w) respectively, utilize the new value of equation (15) calculating w.
These steps constitute the algorithm of one-period, and can repeat this cycle.To study about the noise fluctuations problem that occurs in the heterogeneous networks measured value discussed above now.Although be not have under the situation of noise item, draw this algorithm of cost function, also can be with this algorithm application in the noise cost function.
The effect that repeats above-mentioned algorithm is to make noise contributions and parameter convergence reach mean value.For example, Kushner, H.J.and Clark, D.S. (1978), Stochastic ApproximationMethods for Constrained and Unconstrained System, volume 26 ofApplied Mathematical Sciences, Springer-Verlag, New York, Heidelberg, the random optimization aspect of Berlin is scrutinized this algorithm.By w being increased at any time and reducing, also help to make noise contributions to reach mean value.In addition, in real network, because usually at δ t time cycle integrates measured value, so reduce noise contributions.For just optimised parameter, can select correct δ t, and in optimizing process, can change the value of δ t.
In addition, this algorithm how the variation of the optimum of tracking network be apparent.Even when parameter reaches optimum, this algorithm still causes this point to fluctuate slightly.As long as this optimum does not change, then near this optimum, the mean value that can make this fluctuation is 0.If this optimum value changes, then this algorithm still can be followed the tracks of this variation.
In first embodiment, adjust the configuration parameter value of KPI, recomputate the cost function, this cost function and cost function based on the preceding value of configuration parameter are compared, the value of newly adjusting is used as new configuration parameter, and, adjusts the value of configuration parameter in two steps according to second embodiment.At first, increase the value of configuration parameter, recomputate the cost function, the previous result of this result and cost function is compared according to new value.Then, reduce the value of configuration parameter, recomputate the cost function, the previous result of result and cost function is compared according to new value.Yet,, still can make the variation of configuration parameter small or be 0 even the result of twice previous cost function that changes does not improve.
As a kind of selection, in the 4th embodiment according to the 3rd embodiment, describe cost function and optimization thereof according to a particular network parameter, that is, the particular problem that obtains and optimize cost function according to key performance indicators (KPI), blocked call than (BKCR) is discussed now.
The WCDMA radio interface of 3g mobile network can carry voice service and the data, services with various data rates, business need and quality of service goals.In addition, to big macrocell, very large variation takes place in operational environment from indoor cell.Under various conditions, effectively use the limited frequency band requirement that various important networks and cell parameter carefully are set.The parameter setting is called radio net planning and optimization.Also established the WCDMA network in case set up, its operation and safeguard and mainly be monitoring performance characteristics and mass property and change parameter value in order to improve performance.The Automatic parameter control device is simple, but the performance indicia that it needs index to determine perhaps in this case, need inform clearly that performance is at improvement or the cost function that is worsening.
The target of optimizing is that the total amount with blocked call on the network is reduced to minimum.The special parameter of optimizing is that the soft-handoff parameters window adds (add) (wadd)." Soft handover gains ina fast power controlled WCDMA uplink " Sipila, K.; Jasberg, M.; Laiho-Steffens, J.; Wacker, A.Vehicular Technology Conference, 1999 IEEE 49
Th, Volume:2,1999 Page (s): 1594-1598, vol.2. discusses improving performance according to soft handover.Have been found that when determining to be reduced to minimum cost function, very careful.May produce the cost function that any parameter of selecting all is maintained fixed so that the wrong way combination is every.
By adjusting network parameter, can directly control some factors that influence network performance.For example, number of users, user distribution and type of service.The variation of these external parameters causes the cost function change.This means, be used for cost function is reduced to that minimum any optimized Algorithm all should robust, even and when having random fluctuation, still can realize convergence.Kushner and Clark, " Stochastic Approximation Methods forConstrained and Unconstrained System ", Springer-Verlag, New York, Heidelberg, 1978 stochastic approximation aspect is scrutinized this algorithm.At this, the relevant result who draws about these researchs of cost function optimization problem is even under noise circumstance, when repeating to make intrasystem noise fluctuations reach mean value, optimization problem can be regarded as noiseless optimization.At this, adopt this method, and obtain optimized Algorithm, to optimize definite cost function, in fact, utilize this optimized Algorithm that the noise cost function is reduced to minimum.
Second consideration of the optimized Algorithm of alternative costs function is to change the best operating point of cost function.Therefore, optimized Algorithm should be able to be followed the tracks of any variation of cost function state.By analyzing as can be seen, to compare with the linear convergence of normal gradients algorithm, the algorithm of suggestion has the quadratic convergence that cost function is reduced to minimum.
Quality manager is the logical block that is used to gather the statistics of various performance indicia in the radio network controller.In specified time interval, quality manager calculates these statistics, and this specified time interval is called qminterval.The more operable statistics of quality manager comprise:
● at each qminterval at interval, all connections of this sector of quality manager scrutiny are also checked call quality.In control cycle, quantity that the accumulative total poor quality is called out in two counters and calling sum.Utilize the ratio of Counter Value to obtain quality.
● the blocked call in during the previous qminterval and total ratio that inserts request.
● in formerly during the qminterval, the calling that call drop finishes and the ratio of terminated call sum.
At this, only use blockage ratio, yet, can expand this method and simulation process other statistics to comprise that quality manager returns.
Add according to handoff parameter, the window that is represented as wadd, research will be reduced to the ordinary circumstance of minimum cost function C (with reference to equation 11).If wadd
0Be that C is reduced to the value that minimum window adds.Utilization is calculated C (wadd about the taylor series expansion of wadd
0), draw:
Wherein C ' is the single order differential of C to wadd, and C " be second-order differential.Because C (wadd
0) be minimum, about wadd
0, ask the differential of equation (1), and establish the result and equal 0, then draw,
It is classical Gauss-newton (Gauss-Newton) algorithm of restraining fast.At wadd, must know C ' and C " value.How explanation is by network-evaluated these values now.Note that equation (20) and (21) corresponding to equation (14) and (15), equation (14) is relevant with the updating currently form of described equation with (15).
The research window adds variation from δ wadd to wadd+ δ wadd and the analog value of cost function C (wadd+ δ wadd).Equally, research window be added to the variation of wadd-δ wadd and the analog value C of cost function (wadd-δ wadd), then, utilize algebraic manipulation can with C ' (wadd) and C " (wadd) be expressed as:
If C is a secondary, then therefore there is step convergence.If C is not a secondary, then the expression formula of equation (22) is similar to, but still has the quick convergence faster than normal gradients algorithm.More near wadd
0, just accurate more to being similar to of secondary.Under actual conditions, utilize following process to realize this algorithm:
1) at time t1, the value that window adds is wadd, and according to the network measure value, the value of the function C that assesses the cost (t1).At time t1, the value of wadd is changed to wadd+ δ wadd.
2) time t2 (>t1), the value that window adds is wadd+ δ wadd, directly according to the network measure value, functional value C (t2) assesses the cost.At time t2, the value that window is added changes to wadd-δ wadd.
3) at time t3, the value that window adds is wadd-δ wadd, and directly according to the network measure value, functional value C (t3) assesses the cost.At time t3, utilize equation (20) and equation (21) and (22) to upgrade the value of wadd, wherein
C(wadd)=C(t1)
C(wadd+δwadd)=C(t2)
C(wadd-δwadd)=C(t3) (23)
Repeat this process, thereby cost function is reduced to minimum.It should be noted that by alternately increasing or reducing the value that window adds, can realize two targets.As mentioned above, first target is to estimate C ' and C ".Second target is more no problem.Algorithmic statement is wherein studied to the situation of the minimum value of cost function.At this moment, the gradient of function is 0, and has finished optimization.Yet the optimum of network changes all the time, so cost function also changes all the time.As mentioned above, by replacing the value of wadd, utilize this algorithm can detect and follow the tracks of any this change.
The next stage of the 4th embodiment is that exploitation can utilize above-mentioned optimized Algorithm that it is reduced to minimum cost function.First kind the most generally speaking, can utilize equation (11) to describe cost function, wherein KPI
iBe i KPI of network, and F
iIt is certain function that will define.Every of cost function should be positive all the time, so cost function just is all the time.In addition, function F
iShould not dominate the mode convergent-divergent KPI of cost function with this item in normal operation
iFor example, for the KPI that in claimed range, works, can guarantee correct service quality
iValue, F
i(KPI
i) should be in [0,1] scope.
In the 4th embodiment, only interested in the blocking rate.Purpose is that the blocked call of the function that will add as window is reduced to minimum than (BKCR).In this case, be used to be reduced to minimum explicit costs function be the simple of blockage ratio and.Yet because several reasons, in this case, the better selection of cost function is
C=ulBKCR
2+dlBKCR
2 (24)
Wherein ulBKCR is a up link blocked call ratio, and dlBKCR is a down link blocked call ratio.Yet in any real network, for the acceptable grade of service, the blocked call ratio must be lower than particular value, for example, and 5%.Can further adjust this cost function, so that plugging value is significantly higher than this value.For example,
C=f(ulBKCR)
2+f(dlBKCR)
2 (25)
Wherein
f(x)=exp(x*12)-1 (26)
Select this function to mean, the up link for 5% is blocked, value=1.0 of f (ulBKCR).For the value that is lower than 5%, this function almost is linear.Yet for greater than 5% value, this function increases with index.When having many in cost function, the purposes of function f is more obvious.Another key property of function f is that it is continuously a differentiation function, therefore when asking the derivative of the cost function that is used for optimized Algorithm, does not have problems.
In the 5th embodiment, can expand the algorithm of the 3rd embodiment, in the time that several network parameter will be optimized, cost function is reduced to minimum.By being the problem of the parameter of the 3rd embodiment, realize the situation of a plurality of parameters with a plurality of parameter predigestings.To N the parameter w that will optimize
iVector W study,
W=(w
1;w
2,...;w
N) (27)
Initial value to these parameters of beginning to be optimized from it is studied.Can select this initial value at random,
W
0=(w
0;1;w
0;2;...;w
0;N) (28)
To contain initial value W
0N dimension parameter space in delegation be defined as:
L
0=W
0+λ n
0 (29)
N n dimensional vector n n
0Be unit vector, this unit vector at first has any direction again, and factor λ is a scalar variable.This theory is to follow L
0Cost function is reduced to minimum.This is equivalent to determine the optimum value of λ, and λ is a scalar value, therefore the algorithm that can adopt previous trifle to describe.Suppose that optimum value is λ
0, the equation below then utilizing obtains the new value of W:
W
1=W
0+λ
0n
0 (30)
With another row L
1Be defined as:
L
1=W
1+λ n
1 (31)
N wherein
1Be n
0Conjugate direction.Again, utilize the algorithm of the 3rd embodiment, along the row of redetermination, the repeated optimization cost function.In theory, in noise free system, must be along N conjugate direction (n
0, n
1..., n
N-1), repeat the optimization that follows for N time.For noisy cost function,, must repeat more a plurality of cycles in order to remove noise contributions.There are many well-known methods to produce conjugate direction in each step of optimizing process.Compare with independent optimization network parameter, optimize several network parameters simultaneously, can better cost function be reduced to minimum.
Particularly when this algorithm being expanded to more higher-dimension, adopt the further advantage of this algorithm to be, small by the fluctuation that makes each parameter, can break away from the local minimum value of cost function.
According to the optimization method of the first, second, third or the 4th embodiment not only based on latter two result of cost function, and based on the previous history of cost function measured value.Therefore, at time t, the variation that influences parameter can be cost function and different time t, t-1, t-2, t-3 ... the function of each parameter value of t-n.Therefore, can upgrade each parameter, perhaps be the function of previous measured value with each parametric representation.