Prospects for the use of artificial intelligence in real-time network traffic management

7
163 Prospects for the Use of Artificial Intelligence in Real-Time Network Traffic Management Bob WARFIELD and Peter SEMBER Telecom Australia Research Laboratories, Clayton, Australia Abstract. This paper describes results and plans for further work in the development of a proposed real-time network management system using a combination of conventional opti- misation techniques and artificial intelligence (AI). The objec- tive of the proposed system is to achieve the best possible performance from a circuit-switched network under adverse conditions (particularly traffic overload or equipment failure). It is proposed to develop a real-time AI system which will suggest good control actions quickly, and which will use a combination of conventional optimisation and AI to fine-tune the controls. The design of the proposed system, and its present state of development are described. Keywords. Artificial intelligence, expert systems, network traffic management. Bob Wartield received the degrees of B.E. in 1970 and Ph.D. in 1980 from the University of New South Wales in Sydney. His Ph.D. thesis dealt with problems of teletraffic measurement and forecasting. He currently holds the position of Section Head, Network Analysis in Telecom Australia Research Labora- tories. During 1990 he will be working as a visitor at Bellcore in New Jersey, USA. He has been with Telecom Australia since 1967, starting as a Cadet Engineer in Sydney, where he worked in several positions in the area of network planning and traffic engineering. In 1977 he moved to Melbourne to work in Telecom Australia Headquarters, and later in the Telecom Research Laboratories. His projects have included: performance analysis of packet switched networks, processor capacity studies, dynamic rout- ing, network traffic management, and network architecture. North-Holland Computer Networks and ISDN Systems 20 (1990) 163-169 1. Introduction The problem addressed in this paper is real-time network traffic management in a circuit-switched network--typically an Integrated Digital Network (IDN). In particular, we are concerned with the real-time control of call admission and routing to achieve the best possible performance from a net- work under adverse conditions. Traffic overloads and equipment failure are typical of the condi- tions under which performance needs to be opti- mised. The present method of real-time network traffic management relies on expert human operators who activate network management controls. Usually, they are guided by pre-planned strategies. How- ever, they must cope with whatever network prob- lems arise using expertise and judgement. They require extensive training to equip them to manage even "normal" adverse conditions. In rapidly changing situations or in unfamiliar situations, even expert operators may find that their task is extremely demanding. The objective of the work described here is to assist operators to control the network quickly and effectively. It is planned to achieve this objec- tive by developing a computer system that will advise operators in real-time of suitable control actions. The computer system will employ both AI technology and conventional optimisation tech- Peter Sember received the degree of B.Sc. (Hons) in 1985 and is currently completing a Ph.D. at Monash Uni- versity in Melbourne. His Ph.D. thesis is concerned with the problems of gen- erating intuitive explanations of Baye- sian inferences. He is a member of the AAAI, ACM and IEEE. Presently he holds a position in Network Analysis of Telecom Australia Research Laboratories. He has been with Telecom Australia since 1989. His projects have included the design of a multiprocessor kernel, switched network simulator and of an intelhgent help system for a network designer's workstation. 0169-7552/90/$03.50 © 1990 - Elsevier Science Publishers B.V. (North-Holland)

Transcript of Prospects for the use of artificial intelligence in real-time network traffic management

Page 1: Prospects for the use of artificial intelligence in real-time network traffic management

163

Prospects for the Use of Artificial Intelligence in Real-Time Network Traffic Management

Bob W A R F I E L D and Peter SEMBER Telecom Australia Research Laboratories, Clayton, Australia

Abstract. This paper describes results and plans for further work in the development of a proposed real-time network management system using a combination of conventional opti- misation techniques and artificial intelligence (AI). The objec- tive of the proposed system is to achieve the best possible performance from a circuit-switched network under adverse conditions (particularly traffic overload or equipment failure). It is proposed to develop a real-time AI system which will suggest good control actions quickly, and which will use a combination of conventional optimisation and AI to fine-tune the controls. The design of the proposed system, and its present state of development are described.

Keywords. Artificial intelligence, expert systems, network traffic management.

Bob Wartield received the degrees of B.E. in 1970 and Ph.D. in 1980 from the University of New South Wales in Sydney. His Ph.D. thesis dealt with problems of teletraffic measurement and forecasting.

He currently holds the position of Section Head, Network Analysis in Telecom Australia Research Labora- tories. During 1990 he will be working as a visitor at Bellcore in New Jersey, USA.

He has been with Telecom Australia since 1967, starting as a Cadet Engineer in Sydney, where he worked in several positions in the area of network planning and traffic engineering. In 1977 he moved to Melbourne to work in Telecom Australia Headquarters, and later in the Telecom Research Laboratories.

His projects have included: performance analysis of packet switched networks, processor capacity studies, dynamic rout- ing, network traffic management, and network architecture.

North-Holland Computer Networks and ISDN Systems 20 (1990) 163-169

1. Introduction

The problem addressed in this paper is real-time network traffic management in a circuit-switched network--typically an Integrated Digital Network (IDN). In particular, we are concerned with the real-time control of call admission and routing to achieve the best possible performance from a net- work under adverse conditions. Traffic overloads and equipment failure are typical of the condi- tions under which performance needs to be opti- mised.

The present method of real-time network traffic management relies on expert human operators who activate network management controls. Usually, they are guided by pre-planned strategies. How- ever, they must cope with whatever network prob- lems arise using expertise and judgement. They require extensive training to equip them to manage even "normal" adverse conditions. In rapidly changing situations or in unfamiliar situations, even expert operators may find that their task is extremely demanding.

The objective of the work described here is to assist operators to control the network quickly and effectively. It is planned to achieve this objec- tive by developing a computer system that will advise operators in real-time of suitable control actions. The computer system will employ both AI technology and conventional optimisation tech-

Peter Sember received the degree of B.Sc. (Hons) in 1985 and is currently completing a Ph.D. at Monash Uni- versity in Melbourne. His Ph.D. thesis is concerned with the problems of gen- erating intuitive explanations of Baye- sian inferences. He is a member of the AAAI, ACM and IEEE.

Presently he holds a position in N e t w o r k Analys is of Te lecom Australia Research Laboratories. He has been with Telecom Australia since 1989.

His projects have included the design of a multiprocessor kernel, switched network simulator and of an intelhgent help system for a network designer's workstation.

0169-7552/90/$03.50 © 1990 - Elsevier Science Publishers B.V. (North-Holland)

Page 2: Prospects for the use of artificial intelligence in real-time network traffic management

164 B. Warfiela~ P. Sember / Use orAl in real-time network traffic management

niques. The two technologies offer complementary features. Optimisation of a simple model of the network provides a precise theoretical upper bound on the performance of the network under any conditions. Even new situations that are outside the realm of experience may be modelled and optimised. It is proposed that the AI part of the system will advise the operator of acceptable con- trol actions quickly, and then, over a period of a few minutes, will offer advice on how to fine-tune the controls to achieve the optimal flows.

2. Design of a Real-Time Network Traffic Mana- gement System

An experimental system for real-time control of traffic flows in the IDN is under development. The overall design of the ultimate system is de- scribed in this section, and the main modules are described in more detail in the following sections.

Figure 1 shows the overall design of a system for real-time network traffic management. Under the supervision of a (human) operator, Controls are effected in the Network to improve its perfor- mance. The Actual Performance of the network is measured in terms of an Optimisation Criterion that is defined by the operator. For example, during an emergency such as a fire or an earth- quake, the operator may define the optimisation criterion to be the number of fully routed calls from the emergency area to all other destinations. In quite different circumstances, during a general

I Operator t Supervision Controls

Controls

Actual Performance

Simulated Performance

Target Flows Fine-Tuning L and Data

Module F Optimisation Criterion

(from Operator)

"Good, quick" Control Recommendations

m

N e t w o r k It S i m u l a t o r

Optimal Performance

Optimisation Module

AI Module

Fig. 1. Design of system.

traffic overload, the operator may nominate total revenue as the optimisation criterion.

The objective of the real-time network traffic management system is to recommend the control actions that will achieve the best possible perfor- mance from the network. Achieving the best possi- ble performance will, in general, involve a com- promise between speed and quality of control actions. Given time to perform the necessary com- putations, the controls that will give the optimal performance might be found, but meanwhile con- ditions in the network may have altered. On the other hand, any system that is required to recom- mend control actions immediately can not take into account all possibilities for fine-tuning the performance of the whole network.

A real-time Network Simulator operates in parallel with the network. The simulator allows for experimentation before applying controls to the real network, and also provides estimates of quantities that can not be measured directly in the network. The Simulated Performance provides a useful reference point between the theoretical op- timal and the actual performance.

Data from both the simulator and the network is available to an Optimisation Module which com- putes the optimal pattern of flows in the network relative to the optimisation criterion specified by the operator. This module also computes the theo- retical upper bound to the performance that can be achieved--the Optimal Performance--provid- ing a yardstick against which to rate the simulated and actual performance.

In order to keep the network operating in a reasonably satisfactory way at all times, a real-time AI module is used to compute "good" controls quickly. This module uses the expertise obtained from network operators and coded into Knowledge Areas, as explained in Section 5.

The "good, quick" control recommendations are fine-tuned by a Fine-tuning Module which compares the optimal flows to the actual flows in both the real network and the simulator. The optimal flows are treated as a target, and the controls are adjusted to bring the actual flows into line with the target. Because of the inevitable delay in computing the optimal flows and the appropriate controls, the fine-tuning model must determine when to use the controls recommended by the AI Module, and when to adjust those controls to better match the target flows.

Page 3: Prospects for the use of artificial intelligence in real-time network traffic management

B. Warfield, P. Sember / Use o f A l in real-t ime ne twork traffic managemen t 165

The operator steers and oversees the system. He or she defines the criterion to be used to measure the performance of the network and to compute the optimal performance. Control recommenda- tions are displayed to the operator, who supervises the application of controls to the network. De- pending on the performance of the automatic sys- tem, and on the nature of the situation, the oper- ator may choose to intervene in the application of controls, or merely verify their application. The operator retains the option to over-ride the auto- matic system at any time.

3. O p t i m i s a t i o n M o d u l e

Using a model of the network, the flow pat- terns which optimise a specified objective function are computed. Because of the need to compute optimal flows in real-time (meaning that the re- sults of the computation must be available within minutes of receiving the input data), simple mod- els should be preferred over more accurate but more complex ones. Some experience has been gained in trims of a Linear Program (LP) model which can optimise either revenue or service. Earlier work is reported in [1] and [2], and is briefly summarised in this Section.

Many authors have described designs for dy- namic routing in circuit switched networks. Both feedforward (for example, [3-5]), and adaptive (for example, [6-12]) systems have been described. Filipiak [12] gives an overview of the models that have been applied, and Be1 and others [13] survey dynamic routing systems, classifying them by their structure.

During severe overloads, improved throughput, service, and revenue can be provided by selectively blocking some streams of call attempts. (See, for example [14-16].) Dimensioning of overloaded networks using Linear Programming has been studied in [17]. Linear Programming has also been used by Ash and others in [4] for demand servic- ing.

Some broad assumptions are made about the network to be controlled. The network consists of a number of nodes, each of which can originate traffic, terminate traffic, and switch transit traffic. The nodes are connected by circuit groups that can be one-way or both-way. The set of allowable routes for each stream is specified by the user.

Each circuit group is modelled as a state limiter. It will reject no traffic until its average occupancy reaches its limit, then it will reject all extra traffic. Note that, in this approximate model, there is no random blocking due to congestion. All loss of call attempts is assumed to be selective on the basis of the origin and destination of the call attempt, and the route it is following. Every node in the network is modelled as a rate limiter. This means that it will accept all offered call attempts provided the rate of these attempts is below a fixed limit.

The answer-seizure ratio (ASR) of each destina- tion is assumed to be a measured probability which is independent of the traffic offered to it from the network being managed.

Repeat attempt probabilities for various streams of unsuccessful attempts are modelled as known values.

The following notation will be used: - x d is the rate of flow of attempts from origin i ,j ,k

i to destination j with disposition d along route k. The disposition indicates the type of attempt referred to, which may be any one of

- a t t : call attempts admitted to the network, whether first attempts or repeat attempts,

- r e j : attempts rejected by control action, - n a : call attempts which are not answered,

although fully routed through the network, - s u c : successful calls,

- h~ uc is the holding time of successful calls at destination j ,

- h7 a is the holding times of call attempts which are fully routed to destination j but are not answered,

- h~ .... is the conversation time for successful calls at destination j ,

- h~ et is the setup time for successful calls at destination j ,

- p j is the ASR at destination j - - t h a t is, the conditional probability of a call attempt being answered given that it is fully routed through the network to destination j,

- t,.j is the tariff rate (in dollars per unit of conversation time) for a successful call from i to j ,

- qi,j is the probability of an attempt from origin i to j being repeated,

- R e is the rate limit of the node identified by e, - S e is the state limit of the link identified by e, - q¢.j.k is the indicator variable which takes the

Page 4: Prospects for the use of artificial intelligence in real-time network traffic management

166 B. Warfiela~ P. Sember / Use of AI in real-time network traffic management

Repeat Attempts

Call Intents Origin 1

Abandoned J~t Calls Rate

State Limitar Limiters

Call Ailempts

State Limiters f

A~

Flow Control Calls Destination I

~ Rate Routing and State Lirniter State Flow Control Limiters / ~ t Limiters

f

Call Intents Origin 2

Repeat Attempts

Call Attempts

Successful Calls Destination 2

Abandoned Calls

Fig. 2. LP model of a simple network.

value 1 if the stream of attempts from origin i to destination j via route k passes through the network element identified by e. Figure 2 illustrates the model of a very simple

network. Two origins direct call intents towards two destinations. Each origin-destination pair has one route available, plus the dummy route for rejected attempts. Call attempts which are rejected may be repeated with fixed probability.

In [1] it is shown that, provided selectivity of controls is matched to the discernment of the data gathering process, the optimal solution may be constrained to apply all control action at the input to the network and to have no congestion (accord- ing to the approximate model used) in the net- work. In the following it is assumed that these conditions apply. The resulting optimal control will indicate the quotas of call attempts that may be accepted for allowable routes for every origin- destination pair. The responsibility for determin- ing the real-time routing policy which best meets the specified quotas is devolved to routing deci- sion-makers in individual nodes or in regional network management centers.

The equality and inequality constraints of the Linear Program can now be characterised as fol- lows. Every rate limiter in the network introduces an inequality constraint of the form

E ~ei,j,kXiij, k ~ Re . (1) i , j ,k

The state limiters introduce inequality constraints on the carried traffic. All carded traffic in the network is made up of two additive components: successful traffic and unanswered traffic. Thus the inequality constraint for each state limiter is of the form

E 07,j,k(X~,g~,khg"C+xi~,g,~hja) <- se . (2) i , j ,k

The objective function is chosen as

x-~ suc hconv, att rej f = j ,,.j + x , , j .k j (3) i , j

where - c 1 is the (positive) coefficient for revenue per

unit time: it is a dimensionless quantity, - c 2 is the (positive) coefficient for fully routed

attempts: it is expressed in units of dollars per routed attempt. Optimising the performance of the network can

now be approximated by a Linear Program, in which the objective function f is maximised sub- ject to the constraints implied by the limited re- sources in the network and by the traffic demands.

Fast, real-time solution techniques are pre- sently being researched. The aim is to provide a feasible solution that gives very good performance with acceptable computational delay. Heuristic methods for finding a good starting point are being investigated, as well as methods for speed- ing convergence of large, sparse linear programs.

Page 5: Prospects for the use of artificial intelligence in real-time network traffic management

B. Warfield, P. Sember / Use orAl in real-time network traffic management 167

4 . N e t w o r k S i m u l a t o r

The data collected from the simulator is a su- perset of the data available from the real network. The network simulator provides estimates of all the information that is obtained from the real network, and in addition, allows extra information to be estimated or inferred. For example, details of chain flows and network performance can be estimated from a simulator even if they can not be measured directly. Additional quantities such as, for example, the average number of repeat at- tempts per call intent on particular streams can be easily estimated from the simulator. Of course, the accuracy of such estimates depends on the values of repeat attempt probabilities and other parame- ters.

The simulator may also allow the testing of control actions before they are applied to the real network. The present version of the simulator, which has not been optimised for speed, runs considerably faster than real-time on a thirty node network. In future versions of this module, it is hoped that it will be possible to experiment quickly with possible control actions before they are ap- plied.

The present simulator is a Markov chain simu- lator which models customer repeat attempt be- haviour, node and link limitations and flexible routing rules. It has been developed to be highly interactive, and to allow the following real-time control actions: - change to an alternative set of routing tables, - implement temporary alternative routing, - implement protective controls including: call

gapping selectively by destination, or propor- tional control selectively by link or by destina- tion,

- route calls to Recorded Voice Announcement (which may affect repeat attempt probability),

- protect links by setting dynamic reservation levels,

- implement node overload protection.

5 . A I M o d u l e a n d F i n e - T u n i n g M o d u l e

Several authors have reported on applications of AI to network traffic management--see [18] and [19], for example. The basis of the AI Module in the system described here is the Procedural

Reasoning System (PRS), which is a real-time reasoning and planning system. The following brief description of PRS is adapted from [20].

The system is composed of the following major components: Database, Set of Goals, Set of Knowl- edge Areas, and the Inference Engine.

The database represents the current beliefs of the system. Some of the beliefs are initially pro- vided by the system user and typically includes structural information of a relative static form, for example, the network configuration, node rates and etc. Other beliefs are entered into the data- base during the execution of PRS. They include both network status data, and beliefs derived by PRS itself.

Goals represent the desired behaviours of the system, for example, Determine whether through- put can be increased to node X. An action or sequence of actions is said to achieve the goal if their execution results in a behaviour that satisfies the goal description.

Knowledge Areas (KAs) encode sequences of actions, procedural knowledge, which are either used to satisfy a goal or are simply executed when a particular condition holds. A KA is composed of two major sections, an invocation condition and a body. The invocation condition is a logical expres- sion, which needs to be satisfied before the KA is executed. Since elements of this expression can be both goals and facts, PRS can operate in either a goal-directed or data-driven mode. The body of a KA is a graphical representation of the sequence of actions which are executed when the KA is invoked. Execution commences at the start node and continues through the network from node to node. A transition across an arc can only occur if the goal associated with that arc is satisfied. To determine whether a goal is satisfied, the database needs to be searched for the goal, and if that fails, sub-goals are generated in order to prove the validity of the goal. Execution of a KA terminates when either no arcs can be crossed Or when an end node is reached. Typically, the former termination condition results in the goal which invoked the KA being unsatisfied, whereas the later termina- tion condition results in the goal being satisfied.

The inference engine controls the execution of PRS. Based on the database and the goals, it selects the most appropriate KA and executes it.

An example of a KA for network management appears in Fig. 3. It represents the top level steps

Page 6: Prospects for the use of artificial intelligence in real-time network traffic management

168 B. Warfield, P. Sember / Use of AI in real-time network traffic management

Invocation: Node Processor Alarm on Node $N

? Can Throughput be Increased to $N

Z, ? Can Expansive Controls ? Can Protective Controls

be Implemented be Implemented

? Select ? Select Expans~tro l $0 Protect~trol $C

Fig. 3. A KA for network management.

which need to be taken when a node processor alarm is triggered. We first generate a goal to determine if the throughput to the congested node can in fact be increased. If this goal is satisfied, we then proceed to generate goals to determine the appropriate class of control, expansive or protec- tive. Finally a goal is generated to select an ap- propriate control from a particular class of con- trols.

A goal can either be satisfied heuristically, in which case other KAs are activated, or it can be satisfied analytically. The selection of the ap- propriate strategy depends on their availability and speeds. For example, if the analytical ap- proach for determining the possibility of a throughput reduction to a congested node is numerically intensive, yet a result is required in real-time, we may opt for a less accurate, but fast heuristic strategy.

The fine-tuning module is planned for future development. The difference between optimal per- formance, simulated performance, and actual per- formance (when that can be readily measured) will determine the importance of this module. The optimisation module provides a set of chain flows that will achieve optimal performance. There are two reasons why this optimal performance can not be expected from the actual network: the opti- miser uses an idealised model of the network which assumes no congestion provided capacity is not exceeded; and furthermore, the network flows are determined by offered traffics subject to con-

trols based on destination code and links. There is no direct control of chain flows in the network. Heuristic rules need to be developed to relate link-based controls to the achievement of a de- sired pattern of chain flows.

There are many possibilities for future research and experimentation in the application of AI tech- nology to the system described here. Some exam- pies of possible future work include: explanation of control recommendations, natural language di- alog with operator, case-based reasoning, and machine learning.

6. Conclusion

The system described in this paper is presently under development. Development and testing of the Simulator, the Optimisation Module, and the AI Module are scheduled. Plans for development of the Fine-Tuning Module will depend on the results of the evaluation of the initial experimental system. Early trial results of the optimiser re- ported in [2] indicated the possibility of consider- able improvement in performance of the network, provided the real-time computational problems can be overcome.

Acknowledgment

The permission of the Executive General Manager, Telecom Australia Research Laborato- ries, to publish this paper is gratefully acknowl- edged.

References

[1] R.E. Warfield and D.W. McMillan, A linear program model for the automation of network traffic management, IEEE J. Select. Areas Comm. 6, (4) (1988).

[2] R.E. Warfield and D.W. McMillan, Australian trial of a linear program model for network traffic management, in: Prec. ITC-12, Torino, Italy (1988); also reprinted in ATR 22 (2) (1988) 69-75.

[3] G.R. Ash, R.H. Cardwell and R.P. Murray, Design and optimization of networks with dynamic routing, Bell Sys- tem Tech. J. 60 (8) (1981) 1787-1820.

[4] G.R. Ash, A.H. Kafker and K.R. Krishnan, Servicing and real-time control of networks with dynamic routing, Bell System Tech. J. 60 (8) (1981) 1828-1845.

Page 7: Prospects for the use of artificial intelligence in real-time network traffic management

B. Warfield, P. Sember / Use of A l in real-time network traffic management 169

[5] G.R. Ash, A.H. Kafker and K.R. Krishnan, Intercity dynamic routing architecture and feasibility, in: Proc. ITC-IO, Montreal, Canada (1983).

[6] E. Szybicki and A.E. Bean, Advanced traffic routing in local telephone networks: Performance of proposed call routing algorithms, in: Proc. ITC-9, Torremolinos, Spain (1979).

[7] E. Szybicki and M.E. Lavigne, The introduction of an advanced routing system into local digital networks and its impact on the networks economy, reliability, and grade of service, in: Proc. Internat. Switching Symposium, Paris (1979).

[8] E. Szybicki, Adaptive, tariff dependent routing and net- work management in multi-service telecommunications networks, in: Proc. ITC-11, Kyoto (1985).

[9] T.J. Ott and K.R. Krishnan, State-dependent routing of telephone traffic and the use of separable routing schemes, in: Proc. ITC-11, Kyoto (1985).

[10] G.R. Ash, Use of a trunk-status map for real-time DNHR, in: Proc. ITC-11, Kyoto (1985).

[11] S. Dodd, A decomposition method for DNHR network design on dynamic programming, in: Proc. ITC-I1, Kyoto (1985).

[12] J. Filipiak, Optimal adaptation of telephone traffic to changing load conditions, Large Scale Systems 8 (3) (1985) 237-255.

[13] G. Bel, P. Chemouil, J.M. Garcia F. Le Gall and J. Bernoussou, Adaptive traffic routing in telephone net- works, Large Scale Systems 8 (3) (1985) 267-282.

[14] R.G. Addie and R.E. Warfield, Evaluation of two network management techniques: Holding time minimisation and code blocking, Switching and Signalling Branch Paper 89, Telecom Australia Research Laboratories, 1985.

[15] A. Fukuda, Control algorithms for out-of-chain routing in a telephone network, ECL Japan (USA) 67 (8) (1984) 41-50.

[16] D.G. Haenschke, D.A. Kettler and E. Oberer, Network management and congestion in the U.S. telecommunica- tions network, IEEE Trans. Comm. 29 (1981) 376-385.

[17] A. Guerrero and J.R. De Los Mozos, An approach to traffic analysis of chronically overloaded networks, in: Proc. ITC-IO, Montreal, Canada (1983).

[18] S. Guattery and F.J. Villarreal, NEMESYS: An expert system for fighting congestion in the long distance net- work, in: Proc. Expert Systems in Gooernment Symposium, Maclean, VA (1985).

[19] S. Jiminez, The application of expert systems to network traffic management in Telecom Australia, Telecomm. J. Austra. 38 (1) (1988).

[20] M.P. Georgeff and F.F. Ingrand, Research of procedural reasoning systems, Final Report - Phase 1 SRI Project 2851, SRI International, 1988.