Michael_Phelan_PhD_Introduction

Evolving multi-echelon supply chain inventory andordering policies using biologica y in ired grammar

based meta-heuristics

Michael Patrick Phelan, B.Sc., M.Mangt.Sc.

A thesis submi ed to University College Dublin in part ful lmentof the requirements of the degree of Doctor of Philosophy

Michael Smur t Graduate School of Business,

University College Dublin

August

Supervisor: Dr. Séan McGarraghy

Head of School: Professor Ciarán Ó hÓgartaigh

Abstra

Over the last years, the eld of supply chain management has received widespread a en-tion from researchers andpractitioners across a broad range of disciplines. During this time com-panies havemoved from centrally controlled supply chains towards the outsourcing of non-corefunctions, requiring new and innovative approaches to how these supply chains are optimised.

In recent years there has been a growing literature in the area of biologically inspired algo-rithms, particularly genetic algorithms and genetic programming and their applications to sup-ply chain modelling and inventory control optimisation. Due to the rigidity of the genetic algo-rithms approach, it is difficult to change the underlying model logic and consequently difficultto add richness to the supply chain. While the application of genetic programming provides amore exible approach than that provided by genetic algorithms, to date its application has beenlimited to small supply chain modelling problems in relation to optimal inventory policies.

is research introduces Grammatical Evolution, a relatively new biologically inspired algo-rithm in computer science to the eld of supply chain optimisation, employing human readablerules called grammars. ese grammars provide a singlemechanism to describe a variety of com-plex structures and can incorporate the domain knowledge of the practitioner to bias the algo-rithm towards regions of the search space containing be er solutions.

e primary research question of this work asks if grammatical evolution can provide man-agerial insights and cost effective heuristics for supply chain optimisation across a range of real-istic scenarios.

e methodology used in this research is experimental. Given the stochastic nature of sim-ulating supply chain models with stochastic demand; a statistical analysis of several runs is em-ployed to evaluate the cost effectiveness of supply chain ordering policies generated by grammat-ical evolution.

e supply chains modelled incorporate more realistic features including: inventory alloca-tion policies, payment incentives, linear and distribution supply chain structures, fast and slowmoving stochastic demand and capacity constraints on the warehouse and logistics. Using dif-ferent grammars, ordering policies that minimise the costs in centralised supply chains are com-

xxiv

Michael Patrick Phelan Abstract

pared to policies that balance the associated risks and costs across the supply chain partners indecentralised supply chains.

On the experimental evidence obtained across an extensive range of scenarios, this researchdemonstrates the exibility of grammatical evolution as a supply chain optimisation tool and alsoits ability to adapt to theobjectiveof each scenario, delivering cost effectiveorderingpolicies. egrammars incorporating domain knowledge consistently generate the best supply chain orderingpolicies.

Combining this powerful optimisation approach with more realistic models and incorporat-ing their own domain knowledge, practitioners can develop grammars to bias the grammaticalevolution algorithm towards nding be er supply chain ordering policies. e experiments inthis work demonstrate that grammatical evolution can deliver a range of solutions for the sameproblem, enabling practitioners to compare and contrast policies, highlighting questions that im-pact on the underlying supply chain strategy. However, it is le up to the expertise of the supplychain practitioner to analyse the managerial implications of these policies.

xxv

..

C AIntroduce the topic of this PhD

Identify the research motivation, context, scope and contribution

Present the research questions

Provide an overview of the thesis structure and chapter content

1Introdu ion

. S C M

I since the term Supply Chain Management (SCM) was rst coined, theeld has received widespread a ention from researchers and practitioners alike across a broad

range of disciplines, each a empting to optimise the pro t and minimise the cost to the overallsupply chain. While the traditional stages of supply chain operations: procurement, productionand distribution are as old as commerce itself (Gunasekaran et al., ), the concept of supplychain management represents the “most advanced state in the evolutionary development” ( omasandGriffin, , Gunasekaran et al., ) of these operations. In this time the industrial worldhas seen unprecedented changes in technology from the internet to so ware solutions that candesign supply chain structures, Enterprise Resource Planning systems and forecasting so wareto name but a few, that strive to improve the efficiency of these supply chains.

As the concept of supply chain management was taking hold companies were traditionallyvertically aligned where a single organisation controlled all levels of their supply chain in a cen-tralised structure. In centralised supply chains, organisations could monitor and share all levelsof information throughout the facilities in its chain promoting the shared goal of minimising thecost of the overall supply chain. But during the s there was a signi cant shi towards out-sourcing manufacturing facilities to lower cost economies initially in Europe and then Asia. Inmargin challenged sectors like the PC industry, companies have outsourced their production

Michael Patrick Phelan . Introduction

requirements to contract manufacturers in lower cost countries and created globalised supplychains where predicting customer demand months in advance is critical to the success of theircompany. e lower cost of outsourcing can also have signi cant downsides as Cisco discoveredin resulting in . billion inventory write off (Narayanan and Raman, ) due to poorcoordination with its contract manufacturers and inability to quickly react to a downturn in de-mand. If “SCM implies the optimisation of the whole process om suppliers to end customers” (Leeet al., ) then to truly realise the cost bene ts achieved by these decentralised supply chainsrequires new approaches to how these supply chains are optimised.

In the supply chain management literature the need for closer collaboration between sup-ply chain partners is widely acknowledged. Some of these collaborative initiatives include Ven-dor Managed Inventory (VMI), the apparel industry’s Quick Response (QR) initiative, and theEfficient Consumer Response (ECR) initiative in the grocery industry (Kim and Ryan, ).Coupled with the technological advances of the last years, the amount of data collected bycompanies and the ability to share information across supply chain organisations, suitable in-centives can be devised to ensure that each member of the supply chain acts in the best interestof the overall supply chain. In many cases it may not be in an individual organisations best inter-est to act in a certain way e.g., carry more inventory to ensure shorter lead times and increasedservice levels. Once the supply chain acts in a collective fashion and compensates the organ-isations’ locally sub-optimal strategy (Narayanan and Raman, ), thus ensuring a globallyoptimal strategy, this can lead to even greater efficiencies in todays globalised supply chains.

While the concept of supply chainmanagement only dates back to the early s, the idea ofcoordinated planning dates back to the s with the seminal paper by Clark and Scarf ( )on optimal inventory policies in a multi echelon linear (serial) system with stochastic demand.

is paper was among the rst in the nascent eld of inventory control which has since devel-oped into a widely researched and appliedmethodology, developingmodels that underpinmostinventory management so ware. For most inventory control problems the underlying supplychain model and assumptions are simpli ed approximations of their real world counterparts.

is is necessary to ensure tractability of the models from an optimisation standpoint and thesemodels tend to be very complex requiring a signi cant level ofmathematical understanding. Dueto their complexity and over simpli cation of the real world supply chain models these tech-niques are rarely if ever used by practitioners in an industrial se ing. ere has been signi cantresearch into more applicable heuristic techniques where the underlying models tend to moreclosely re ect the complexity of real world supply chains and the methods can be more easilyapplied by non-mathematicians. Similarly in the eld of forecasting, it is also themore easily un-derstood and implementable techniques (e.g., Moving Average & Exponential Smoothing) that arewidely used in industry through dedicated so ware packages or spreadsheet modelling.


. B I A & S C

Since themillennium there has been a growing literature in the area of biologically inspired algo-rithms and supply chain optimisation most notably Genetic Algorithms and Genetic Program-ming. e most common type of genetic programming used is John Koza’s (Koza, ) treebased form. Using these approaches arti cial agents act in the place of the supply chain facili-ties or stages evolving supply chain ordering policy heuristics that minimise the total cost of theoverall supply chain. Continuing in this vein, supply chain heuristics are evolved using biologi-cally inspired algorithms utilising existing inventory control and forecasting models that can beapplied in an industrial se ing to a variety of supply chainmodels for this research. is researchintroduces a relatively new evolutionary algorithm in computer science to the eld of supplychain optimisation. Grammatical Evolution (GE) (O’Neill and Ryan, , ) is an evolu-tionary algorithm, a form of linear genetic programming. It can evolve supply chain orderingpolicies similarly to Koza’s genetic programming. One of the main differences between Koza’sgenetic programming and grammatical evolution is that grammatical evolution employs a userde ned grammar to construct the ordering policies that can adhere to rules as laid out by the user.

ese grammars are human readable rules based on the underlying problem, in this case supplychain ordering policies. e user can specify the building blocks for the supply chain orderingpolicy incorporating traditional inventory control and forecastingmodels that will minimise theoverall supply chain cost. Within these grammars the user can also provide guidance to the algo-rithm on how to allocate constrained supply and what incentives if any should be given to agentsin the supply chain.

Sowhat separates grammatical evolution from its genetic programmingandgenetic algorithmcousins particularly in the supply chain domain? Grammars!

Step back from supply chains for a moment and consider a grammatical evolution grammarle containing all the rules required to compose a symphonyor a sonnet. A er running the gram-

matical evolution algorithm, the resulting symphony or sonnet would be syntactically correct asper the rules of the respective medium, but how could they be assessed in order to allow thealgorithm evolve, keeping the best components from previous generations? Would grammati-cal evolution be able to create a symphony more revered than Beethoven’s thor a collection ofsonnets to rival Shakespeare? is would not be possible, as to assess the output from grammat-ical evolutions work due to the number of compositions¹ would be prohibitive and even moreimportantly one person’s Beethoven is another person’s Led Zeppelin!

Perhaps this is neatly summarised by Sir Geoffrey Jeffersons’ Lister Oration in titled“ e Mind of Mechanical Man”:

“Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions

¹ generations with a population of individuals would generate , potential solutions to be as-sessed and scored


felt, and no: by the chance fall of symbols, could we agree that machine equals brain – that is, not onlywrite it but know that it had wri en it.”

In the same year AlanTuring, who is o en creditedwith being the father of computer scienceand the father of arti cial intelligence (Beavers, ) was quoted in the Times as saying:

“... I do not see why it [a computer] should not enter anyone of the elds normally covered by thehuman intellect, and eventually compete on equal terms. I do not think you can even draw the line aboutsonnets, though the comparison is perhaps a li le bit unfair because a sonnet wri en by a machine willbe be er appreciated by another machine.”

While it is far beyond the scopeof thiswork todiscuss the potential for grammatical evolutionto compose sonnets, grammatical evolution is an excellent choice for supply chain optimisation.Although the grammatical evolution algorithm cannot “understand” the supply chain orderingpolicies being generating, it can interpret and score the policies’ efficiency based on commonmetrics such as cost or service level and their level of associationwith the underlying policy, i.e., alower cost policy is more efficient and the converse is also true. is difference is highlighted in § .and stresses the importance, that like any inventory control or forecasting tool, the solutionsprovided by grammatical evolution need to be analysed in amulti-facetedmanner to ensure bothmetric drivers (cost, service level etc.) and organisational strategies are aligned. As mentioned,the power of grammatical evolution lies in its human readable grammars that can incorporatedomain knowledge and provide insights to the underlying problem, in this case supply chainordering strategies.

. R M

emotivation for this research is to develop a theoretical framework to generate heuristic solu-tions for intractable supply chain optimisation problems. Of particular interest is the generationof cost minimising supply chain ordering policies that balance the costs and risks across all part-ners in decentralised supply chain models.

e increasing complexity ofmodern global supply chains, requires innovative heuristic tech-niques that can model more realistic supply chains, including the supply chain structure, cus-tomer demand distributions, costs incurred by supply chain partners and other supply chainmanagement concepts (e.g., allocation policies and payment incentives).

In recent years there has been a growing literature in the area of biologically inspired algo-rithms, particularly genetic algorithms and their application to supply chain optimisation. Dueto the rigidity of the genetic algorithms approach, and the difficulty of incorporating more real-ism into supply chain modelling, other methods are investigated.

Evolutionary automatic programs including grammatical evolution have demonstrated theirefficacy in nancial models where the tness function is complex, requiring the balancing of re-


turn and risk (O’Neill et al., a). Similar complexity can be found in supply chains where thechallenge of minimising costs and maximising customer service levels provide a rich problemdomain for grammatical evolution.

e grammar based approach in grammatical evolution enables the end user to develop tai-lored grammars dependingon their underlying investigation and requirementswithout changingthe model or algorithm. It also allows the user to incorporate domain knowledge, biasing² thegrammatical evolution algorithm towards regions of the search space containing be er orderingpolicies.

is research presents a framework, incorporating grammatical evolution to generate supplychain heuristics that enable supply chainmanagement practitioners to test different supply chainscenarios and policies. It is le up to the expertise of the supply chain practitioner to analyse themanagerial implications of these policies.

. R C

A detailed discussion of the research context behind this work is presented in Chapter , whilea brief overview is provided in this section. e most closely related work to this research inthe area of biologically inspired algorithms and supply chain management is that of Kimbroughet al. ( ), Moore and DeMaagd ( ), Chan et al. ( ), O’Donnell et al. ( ), Luet al. ( ) and Kleinau and onemann ( b). Kimbrough et al. ( , ) use arti cialagents to replace human players in the Beer Game where a genetic algorithm is used to generatethe ordering policies. ese arti cial agents generate policies that reduce the demand ampli -cation between agents in the Beer Game thus mitigating the bullwhip effect. e agents act in amanner that are “(apparently) Nash equilibria” in that they generate the optimal ordering policiesindependent of eachother resulting in a lower cost to the overall supply chain. O’Donnell ( )and O’Donnell et al. ( , ) extend the work of Kimbrough et al. ( ) by incorporat-ing widely used forecasting techniques and examining how the genetic algorithm agents react tosome of the known cases of the bullwhip effect (e.g., price uctuations) while Lu ( ) and Luet al. ( , ) extend these results further with the inclusion of more realistic cost struc-tures into the underlying model. Chan et al. ( ) also build on and benchmark against thework of Kimbrough et al. ( ) by allowing agents employ traditional inventory control poli-cies and report lower overall supply chain costs using this approach. Kleinau ( ) andKleinauand onemann ( b) employ genetic programming to evolve inventory control models ina single and two echelon supply chain using a variety of cost se ings common to the inventorycontrol literature. Whereas Chan et al. ( ) use genetic algorithms to select well known in-

²It is crucial that any optimisation algorithm should be biased towards what is considered to be the optimalsolutions. e word bias in this context has no negative connotations: it simply describes the tendency of thealgorithm to move towards the best regions of the search space.


ventory control models and parameters, the work of Kleinau and onemann ( b) employgenetic programming to discover the structure and parameters of themodels themselves. Mooreand DeMaagd ( , ) also use genetic programming to evolve ordering policies to min-imise the total cost of running the BeerGame, but their work is concernedwith identifying goodmodel parameters for genetic programming as opposed to the supply chain insights gained.

Having presented a brief overview of the literature most closely related to the work in thisthesis, the research gaps identi ed by this work are presented in the next section.

. R G

e broad area of supply chain management provides researchers with a rich set of problemsspanning multiple academic elds, including two conceptually parallel yet interconnected eldsof study in the inventory control and supply chain literature, both dating back to the late s.

e rst related area of study concerns the development of optimal mathematical models intraditional inventory control theory (Clark and Scarf, , Axsäter, a). ese models pro-vide optimal inventory ordering policies to minimise the supply chain costs in linear and distri-bution models where demand and lead times can be deterministic or stochastic. However boththe limited supply chain structures and modelling assumptions are unrealistic and the complexmathematical solutions make them una ractive to practitioners. While exact solutions exist forlinear and distribution supply chain problems, it is common in practice to implement heuristicsto obtain approximate solutions because they are easier to understand and implement (Axsäter,

b).As mentioned in the previous section, Kleinau and onemann ( b) use genetic pro-

gramming to evolve inventory control models and heuristics in single and two echelon supplychains. is work demonstrates the power of evolutionary automatic programs and their abilityto adapt to different supply chain models, but also lacks sufficient realism to be of use in a realworld se ing.

e second area of research into supply chain optimisation follows amodelling approach, de-veloped by Forrester ( ) andmore recently popularised by Lee et al. ( a,b), investigatingthe causes and solutions for the demand ampli cation that occurs in supply chains commonlyreferred to as the bullwhip effect (also known as the Forrester effect). e Beer Game is a popularsimulation tool developed by Jarmain and Fey ( ) used by academics and practitioners aris-ing from the work by Forrester ( ). It is in this research space that Kimbrough et al. ( ),Chan et al. ( ), O’Donnell et al. ( ) and Lu et al. ( ) employ genetic algorithms tomitigate the bullwhip effect using arti cial agents in a Beer Game simulation. While the BeerGame is a useful optimisation tool for modelling simpli ed linear supply chains and is also im-plemented in this research, it lacks sufficient realistic features (i.e., supply chain structure, customer


demand, costs etc.) required to capture the complexity of real supply chains.Examining the implementation of the genetic algorithm approach, and the encoding strat-

egy used by Kimbrough et al. ( ), Chan et al. ( ), O’Donnell et al. ( ) and Lu et al.( ), each potential solution is encoded in a xed length binary string. is approach lim-its the structure and complexity of the (decoded) resulting ordering policy, making it harder tochange the underlying model logic and thus harder to add richness to the supply chain.

Both of these areas use different approaches including optimal mathematical models, heuris-tics and simulation models to optimise supply chain costs in simpli ed supply chain models. Inrecent years different biologically inspired algorithms have been applied to these areas, incorpo-rating more realistic features adding to the complexity and intractability of these supply chainmodels.

While themore recent contributions to the researchdiscussed in this section combinebiolog-ically inspired algorithms and supply chain modelling, their approaches lack sufficient realism.

e genetic algorithms used to mitigate the bullwhip effect employ simple demands and coststructures unlike the genetic programming approach where more realistic demand and costs areused, whereas the genetic programming supply chainmodels are trivial unlike themore complexmodels incorporated into the genetic algorithm approach.

Having identi ed several open research areas in the domain of biologically inspired algo-rithms applied to supply chain management, the next section identi es the topics selected forresearch in this thesis and the reasons for these choices.

. R S

e scope of this research is interdisciplinary, primarily focused on supply chain managementand biologically inspired algorithms, but also incorporating techniques from inventory control,forecasting and simulation without contributing to their body of knowledge. While this mayseem like a broad range of subjects, it is becoming increasingly common in supply chain optimi-sation to span multiple disciplines seeking out new and innovate approaches given the complexchallenges faced by modern global supply chains.

In the previous section, two recent evolutionary algorithm approaches to supply chain mod-elling are discussed. e research in this thesis builds on the work of Kimbrough et al. ( ),Chanet al. ( ),O’Donnell et al. ( ), Luet al. ( ) andKleinauand onemann( b),combining and extending these different approaches with the addition of more realistic supplychain features and supply chain management concepts used in real world supply chains.

Grammatical evolution is another technique in the burgeoning eld of biologically inspiredalgorithms and its application to supply chains provides researchers and practitioners alike witha new tool to tackle supply chain complexity. Employing grammatical evolution to generate the


supply chain ordering policies, the structure of the resulting policy does not need to be knowna priori, allowing the agents more exibility to adapt to more realistic supply chain conditions.Using grammatical evolution, multi-echelon serial and distribution supply chains are simulatedwhere the ordering policies generated by each agent in the supply chain can incorporate domainknowledge from commonly used inventory control and forecastingmodels. ese ordering poli-cies are evolved over time to generate themost cost effective policies for the supply chain and arebenchmarked against popular inventory control heuristics and the aforementioned genetic algo-rithm approaches. Similarly to related inventory control models and the Beer Game simulation,the supply chains modelled for this research are also limited to single product items where re-turns are not permi ed. A list of themodelling assumptions employed for this work is presentedin § . . .

Owing to the interdisciplinary nature of biologically inspired algorithms applied to supplychain management, the following principle is applied to demarcate the scope of this work: (A)research into the causes and potential solutions that mitigate the bullwhip effect are excluded,except where required to benchmark against the grammatical evolution approach and (B) areasof research that extend into additional literature domains beyond those already mentioned arealso outside the scope of this work to avoid diluting the academic contribution of this research.

As mentioned, this research is related to the work of Kimbrough et al. ( ), Chan et al.( ), O’Donnell et al. ( ) and Lu et al. ( ) who examine the causes of the bullwhipeffect (e.g., price& sales uctuations) and present solutions tomitigate against it in a simpli ed lin-ear supply chain i.e., the Beer Game. However the work in this thesis is not explicitly interested inthe causes and potential solutions that mitigate the bullwhip effect. Instead, the objective of thegrammatical evolution arti cial agents is to generate supply chainorderingpolicies thatminimisethe overall supply chain costs, while balancing the associated risks and costs across the agents indecentralised supply chains, incorporating supply chain management concepts including allo-cation policies and payment incentives. e agents are also provided with a suite of popularinventory control and forecasting models to generate efficient ordering policies in more realisticsupply chains. Both linear and distribution supply chain structures are modelled, with slow orfast moving customer demand and different cost structures, including capacity constraints onwarehouse and logistics. ese additional supply chain factors, enable the modelling of morerealistic supply chains in the eld of biologically inspired algorithms and supply chain optimisa-tion. While supply chain costs are the only metric driving the effectiveness of these policies, theservice levels associated with implementing these ordering policies and demand ampli cationbetween echelons are monitored.

Both linear and distribution supply chainmodels are examined in this research, adding to theliterature in the application of biologically inspired algorithms to supply chain optimisation. As-sembly systems as discussed in § . . . are common in process industries (Axsäter, b), re-


quiring downstream customers to order frommultiple suppliers which inmodelling terms couldbe achieved using Game eory and is outside the scope of this research.

Another research area that extends into the eld of Game eory relates to supply chain con-tracts that are discussed in § . . . ese contracts are also outside the scope of this research,however the supply chain agents in this research can employ linear transfer payment schemesbased on inventory and backorders similar to the work of Cachon and Zipkin ( ).

. R Q

Having identi ed research gaps in the area of biologically inspired algorithms applied to sup-ply chain management and the scope of this work, a set of targeted research questions are nowpresented to address the gaps within this scope.

..

P R QCan Grammatical Evolution provide managerial insights and cost effective heuristics forsupply chain optimisation across a range of realistic scenarios?

is primary research question is examined via the grammars and their resulting policies acrossdifferent scenarios:

RQ What is the impact on the ordering strategies using different customer demand distribu-tions to represent fast and slow moving items?

RQ What is the impact of different cost structures (e.g., logistics, excess inventory, inventory hold-ing costs vs. backorder ratios) on the overall performance of the supply chain?

RQ What grammars employ “order-in ating” or “truth inducing” allocationpolicies anddo theyhave any signi cant in uence on the overall supply chain strategy?

RQ What impact can cost incentives have on the overall supply chain performance anddo theyin uence agents ordering strategies?

RQ What is the impact of decentralised agent decisionmaking (i.e., each agent generates a policytominimise their own cost base regardless of other agents) on the overall supply order strategyand how does this impact on total costs?

RQ Can the grammatical evolution arti cial agents generate supply chain ordering policiesthat mitigate the bullwhip effect?

ese research questions can be divided into two categories: ( ) the questions addressing themotivation behind this work (RQ , RQ & RQ ), and ( ) the questions that benchmark this


approach against related approaches and also test the exibility and robustness of the solutionspresented using grammatical evolution (RQ , RQ & RQ )

An extensive set experiments are run to answer the research questions posed in this section.Each experimental group is designed for a speci c purpose and is described in § . . Each re-search question has an associated experimental group but there is also crossover between groupsand research questions.

. A C &M I

e research presented in this thesis is interdisciplinary incorporating techniques from supplychain management, computer science, inventory control, forecasting and simulation but doesnot contribute to the literature of all these areas. e contributions made by this work are in thesupply chainmanagement andcomputer science eldswhereasmethods andmodels are adoptedfrom the other elds without contributing to their body of knowledge.

On the evidence obtained from the extensive set of experiments designed to test the efficacyof grammatical evolution as a supply chain optimisation tool across a range of scenarios, this re-search demonstrates that grammatical evolution is able to adapt to different scenarios and delivercost effective solutions that can also mitigate the bullwhip effect while maintaining a high levelof customer service. A range of scenarios are tested including different cost structures, capacityconstraints on warehousing and logistics, slow and fast moving customer demand across bothlinear and distribution supply chain models.

A secondary minor contribution is in the computer science eld of grammatical evolution asthis research is the rst application of grammatical evolution to the domain of supply chainman-agement (Phelan and McGarraghy, c) adding to grammatical evolution’s set of benchmarkapplications (which heretofore have included nance, environmental and music). Grammatical evo-lution has a modular structure including the choice of search engine that can be used to evolvethe supply chain ordering policies. For this research three engines are examined: Genetic Algo-rithms, Quantum Inspired Genetic Algorithms and Particle Swarm Optimisation to determinethe most efficient engine, also adding to the existing grammatical evolution literature on engineselection.

It is hoped that by adding extra richness into the models, making it more applicable to realworld supply chains, this type of evolutionary supply chain ordering policy heuristic generatorcan be used in a supply chain managerial se ing. One of the key factors making grammaticalevolution an excellent choice as a supply chain heuristic generator is its grammar. Grammati-cal evolution, like other evolutionary automatic programs, is useful when the tness function iscomplex (e.g., balancing of return and risk in nance, balancing pro ts and customer service levels in asupply chain) and there is the ability to incorporate domain knowledge (e.g., inventory control and


forecasting models) that can bias the search space. Grammatical evolution also produces humanreadable rules that have the potential to enhance understanding of the problem domain. Alsothe form of the model does not need to be known a priori, for example, where the modeller hasa good understanding of the variables in the model but may have a weak understanding of therelationships within the model. By adjusting grammars that bias the search space the modellercan gain a be er understanding of the problem domain and evolve their own thought processinto targeting optimal supply chain solutions.

A so ware package called SupplyGEn was wri en in Java to provide additional realism andfunctionality to the supply chain models. is is described in more detail in § . .

. T S

An overview of the thesis structure can be seen in Figure . . where chapters are colour codedto re ect their relationship to each other.

Chapter - Introduction e introduction chapter presents a high level overview of this re-search including itsmotivation, context, scope, researchquestions andcontributions. Someof the problems facingmodern supply chains are introduced, particularly, ordering strate-gies in decentralised supply chains and how to encourages stages in the supply chain to actin the best interest of the overall chain. An overview of biologically inspired algorithmsapplied to supply chain optimisation and mitigation of the bullwhip effect is presented.

Chapter - Literature Review e literature review chapter provides more detail on the con-cepts introduced in Chapter and further describes the gap lled by this research. emotivation for this research is driven by the need to nd innovate solutions to addressthe increased complexity in supply chain management. e various concepts from sup-ply chain management including supply chain metrics, allocation policies and paymentincentives pertaining to this research are described in detail. e bullwhip effect is dis-cussed and the use of the Beer Game as a supply chain simulation model introduces theeld of biologically inspired algorithms to supply chain optimisation.

Chapter - Inventory Control, Forecasting&Benchmarking is chapter describes the in-ventory control and forecasting models that are used in this work, in particular that areavailable to arti cial agents in the SupplyGEn system. An overview of the models usedin this research comprising diagrams and equations is provided including a descriptionof the inventory control heuristics used to benchmark against the grammatical evolutionheuristics in Chapter .


Figure 1.9.1: Thesis structure


Chapter -Methodology e methodology chapter presents a more detailed overview of thebiologically inspired algorithms discussed in Chapter , including a detailed example ofhowgrammatical evolution is used togenerate supply chainorderingpolicies. Anoverviewof the approach to evaluate each ordering policy via a supply chain simulation incorporat-ing agent based modelling is also provided.

Chapter - Experimental Design e experimental design chapter outlines the eleven experi-mental groups comprising of grammatical evolution experiments designed to test for:(A) grammatical evolution parameters se ings and grammars and (B) the various supplychain structures and se ings as discussed in Chapters and . A detailed description ofthe Java model developed to assess the tness of each individual in the grammatical evo-lution model is provided.

Chapter - Results is chapter presents the results of the experiments across the ex-perimental groups. ese experiments aredesigned toanswer the researchquestionsposedin § . and test the various supply chainmodel structures and scenarios while also testingfor good grammatical evolution parameter se ings.

Chapter - Discussion is chapter discusses the results from the experimental groups andhow they answer the research questions in § . . e rst experimental groups are alsoused to determine the best benchmarkmodels for each supply chain structure and to iden-tify the most appropriate simulation output analysis method. e best grammars acrossthe experimental groups are identi ed. e best policies are also examined in terms oftheir service levels and demand ampli cation across the supply chain.

Chapter - Conclusions e conclusions chapter provides a summary of the overall thesis. Italso discusses the signi cance of this work, including the contributions to the academicliterature in this area and to management and practice. e limitations of this work arealso discussed along with suggestions for future research arising from this work.

Appendices e appendices provide additional information to the main body of the thesis in-cluding: list of abbreviations, SupplyGEnmodel le de nitions, Java code functions, sim-ulation output graphs, grammatical evolution best policy occurrence graphs and the SQLScript to analyse results data.

. S

A high level overview of this research including its motivation, context, scope, research ques-tions and contributions is presented in this chapter. Given the complexity of modern globalsupply chains and the need to investigate and develop innovative solutions that can be applied


in industry, in recent years the eld of biologically inspired algorithms has been proposed toaddress these problems. is research introduces grammatical evolution, a relatively new com-puter science technique, to the eld of supply chain optimisation, building on and advancingthe existing work in this area. e grammar based structure of grammatical evolution makes itan ideal choice for supply chain optimisation given their human readability and ease of imple-mentation for practitioners. Coupled with the ability to re ne the search space by incorporatingdomain knowledge, improving the quality of results, these grammars can also expose facets of theunderlying problem, enhancing the user’s understanding of the problem at hand. is researchpresents a theoretical framework, incorporating grammatical evolution to generate supply chainheuristics that enable supply chain management practitioners test different supply chain scenar-ios and policies. However, it is le up to the expertise of the supply chain practitioner to analysethe managerial implications of these policies.