discrete event simulation, system dynamics and agent …

1
Discrete Event Simulation, System Dynamics and Agent Based Simulation: Discussion and Comparison Robert Maidstone STOR-i DTC, Lancaster University Introduction Simulation modelling is an important tool in Operational Research: It provides a method to approximate the behaviour in the real system (and hence can be used for testing scenarios), Also constructing the model can prove useful in achieving greater understanding of the system. Many different simulation techniques are used in Operational Research, here three of the main ones are discussed and compared: Discrete Event Simulation (DES), System Dynamics (SD), Agent Based Simulation (ABS). Each of these techniques has its own merits and failures and when choosing which one to use it’s important that the modeller picks the method which best reflects the system. Discrete Event Simulation Discrete Event Simulation (DES) is probably the most widely used simulation technique in Operational Research. As the name suggests it models a process as a series of discrete events. DES models are built using: Entities - The general name for the objects that move through the system. Events - The processes which the entities pass through. Resources - Objects which are needed to trigger events. Figure 1: A typical DES model (created using SIMUL8) System Dynamics System Dynamics (SD) takes a different approach to DES, focusing on flows around networks rather than queueing systems, it considers: Stocks - basic stores of objects. Flows - define the movement of objects between different stocks in the system. Delays - delays between the measuring and then acting on that measurement. Figure 2: A typical SD model (created using Vensim) Agent Based Simulation ABS is a relatively new technique to be used in OR. ABS consists of: Autonomous Agents - These are self-directed objects which move about the system, Rules - which the agents follow to achieve their objectives. Agents then move about the system interacting with each other and the environment. Can be used to model many situations in which the entities have some crude intelligence Can often produce “nice” graphics and animations which highlight the behaviour of the system, such as in Figure 3. Figure 3: Variation patterns of Conway’s Game of Life (Chan et al., 2010) Comparing DES and SD models DES and SD both have advantages and disadvantages and work well when modelling different systems. Even when both are used to model a single system it may be the case that both are still useful, as they can lead to different conclusions and insights about the system. The table below gives a few of the key differences between the two approaches. DES SD Often used to model situa- tions which are (or can be approximated by) networks of queues. Used to model situations which form flows or larger sys- tems where flows are a good approximation. Discrete Continuous Often thought of as micro- scopic Often thought of as macro- scopic Stochastic - and uses proba- bility distributions Usually deterministic In addition a couple of empirical studies of the differences between DES and SD have been carried out: The first (Tako and Robinson, 2009a) concluded that as far as the managers using the model were concerned there was statistically no difference between which of the two methods was found to be easier to use. The second (Tako and Robinson, 2009b) looked at the differences in the actual modelling procedures. It was found that in general DES modellers spent more time modelling,verifying and validating, and SD modellers spent more time in the conceptual modelling stage. Combining DES and SD Many situations can be modelled as a combination of DES and SD (in a hybrid type model), hence making full use of the advantages of each technique. True hybrid models have never really been constructed, this is mainly due to limitations in the software available. However many models have been constructed in which a DES model and a SD model communicate (through a third piece of software). One situation where combining models is often used, is when external influences are needed to be modelled without adding too much detail. Here an SD model can be used to deal with the wider system, whilst a DES model can look at the key areas. Comparisons with ABS As with SD and DES, ABS again offers a slightly different tool for modelling. Comparisons with DES: Agents in ABS have their own goals and behaviour (active), behaviour of entities in DES models is determined by the system (passive). In DES queues are a key element, whereas there is no concept of queues in a ABS system. Both are stochastic in nature and can involve input distributions to model random behaviour. Often DES/ABS models are used, which consist of a DES system with some active entities added to it. Comparisons with SD: It can be shown that in fact every well formulated SD model has an equivalent formulation as an ABS model. This is despite the fact the SD is deterministic in nature, compared to ABS’s stochasticity. ABS has often been looked over when it comes to modelling systems this is for a number of reasons: Lack of easy to use software. High amount of time needed to develop model (means that it’s less effective to find a “quick and dirty” solution to a problem) Reluctance from OR practitioners to move away from more established techniques. Conclusions DES, SD and ABS all have benefits and disadvantages and are applicable in different situations. A problem in simulation modelling is a tendency for a system to be modelled using the technique which the modeller feels most comfortable with. Instead the problem should determine the method used to model it. In order to choose the right method knowledge of a range of techniques is needed as well as detailed knowledge of the system and the objectives of the simulation. http://www.lancs.ac.uk/ ~ maidston [email protected]

Upload: others

Post on 10-Nov-2021

14 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Discrete Event Simulation, System Dynamics and Agent …

Discrete Event Simulation, System Dynamics andAgent Based Simulation: Discussion and Comparison

Robert MaidstoneSTOR-i DTC, Lancaster University

Introduction

Simulation modelling is an important tool inOperational Research:

It provides a method to approximate thebehaviour in the real system (and hence can beused for testing scenarios),

Also constructing the model can prove useful inachieving greater understanding of the system.

Many different simulation techniques are used inOperational Research, here three of the main onesare discussed and compared:

Discrete Event Simulation (DES),

System Dynamics (SD),

Agent Based Simulation (ABS).

Each of these techniques has its own merits andfailures and when choosing which one to use it’simportant that the modeller picks the method whichbest reflects the system.

Discrete Event Simulation

Discrete Event Simulation (DES) is probably themost widely used simulation technique inOperational Research. As the name suggests itmodels a process as a series of discrete events. DESmodels are built using:

Entities - The general name for the objects thatmove through the system.

Events - The processes which the entities passthrough.

Resources - Objects which are needed to triggerevents.

Figure 1: A typical DES model (created using SIMUL8)

System Dynamics

System Dynamics (SD) takes a different approach to DES,focusing on flows around networks rather than queueingsystems, it considers:

Stocks - basic stores of objects.

Flows - define the movement of objects between differentstocks in the system.

Delays - delays between the measuring and then acting onthat measurement.

Figure 2: A typical SD model (created using Vensim)

Agent Based Simulation

ABS is a relatively new technique to be used in OR. ABSconsists of:

Autonomous Agents - These are self-directed objects whichmove about the system,

Rules - which the agents follow to achieve their objectives.

Agents then move about the system interacting with eachother and the environment.

Can be used to model many situations in which the entitieshave some crude intelligence

Can often produce “nice” graphics and animations whichhighlight the behaviour of the system, such as in Figure 3.

Figure 3: Variation patterns of Conway’s Game of Life (Chan et al., 2010)

Comparing DES and SD models

DES and SD both have advantages and disadvantages andwork well when modelling different systems. Even when bothare used to model a single system it may be the case thatboth are still useful, as they can lead to different conclusionsand insights about the system. The table below gives a few ofthe key differences between the two approaches.

DES SDOften used to model situa-tions which are (or can beapproximated by) networks ofqueues.

Used to model situationswhich form flows or larger sys-tems where flows are a goodapproximation.

Discrete ContinuousOften thought of as micro-scopic

Often thought of as macro-scopic

Stochastic - and uses proba-bility distributions

Usually deterministic

In addition a couple of empirical studies of the differencesbetween DES and SD have been carried out:

The first (Tako and Robinson, 2009a) concluded that as faras the managers using the model were concerned there wasstatistically no difference between which of the twomethods was found to be easier to use.

The second (Tako and Robinson, 2009b) looked at thedifferences in the actual modelling procedures. It was foundthat in general DES modellers spent more timemodelling,verifying and validating, and SD modellers spentmore time in the conceptual modelling stage.

Combining DES and SD

Many situations can be modelled as a combination of DESand SD (in a hybrid type model), hence making full use of theadvantages of each technique.

True hybrid models have never really been constructed, thisis mainly due to limitations in the software available.

However many models have been constructed in which aDES model and a SD model communicate (through a thirdpiece of software).

One situation where combining models is often used, iswhen external influences are needed to be modelled withoutadding too much detail. Here an SD model can be used todeal with the wider system, whilst a DES model can look atthe key areas.

Comparisons with ABS

As with SD and DES, ABS again offers a slightlydifferent tool for modelling.

Comparisons with DES:

Agents in ABS have their own goals andbehaviour (active), behaviour of entities in DESmodels is determined by the system (passive).

In DES queues are a key element, whereas there isno concept of queues in a ABS system.

Both are stochastic in nature and can involveinput distributions to model random behaviour.

Often DES/ABS models are used, which consistof a DES system with some active entities addedto it.

Comparisons with SD:

It can be shown that in fact every well formulatedSD model has an equivalent formulation as anABS model.

This is despite the fact the SD is deterministic innature, compared to ABS’s stochasticity.

ABS has often been looked over when it comes tomodelling systems this is for a number of reasons:

Lack of easy to use software.

High amount of time needed to develop model(means that it’s less effective to find a “quick anddirty” solution to a problem)

Reluctance from OR practitioners to move awayfrom more established techniques.

Conclusions

DES, SD and ABS all have benefits anddisadvantages and are applicable in differentsituations.

A problem in simulation modelling is a tendencyfor a system to be modelled using the techniquewhich the modeller feels most comfortable with.

Instead the problem should determine the methodused to model it.

In order to choose the right method knowledge ofa range of techniques is needed as well as detailedknowledge of the system and the objectives of thesimulation.

http://www.lancs.ac.uk/~maidston [email protected]