Clustering Project Risks According to Their Interactions

8th International Conference of Modeling and Simulation - MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia “Evaluation and optimization of innovative production systems of goods and services” CLUSTERING PROJECT RISKS ACCORDING TO THEIR INTERACTIONS Ludovic-Alexandre VIDAL, Franck MARLE Ecole Centrale Paris Laboratoire Genie Industriel Grande Voie des Vignes 92290 Châtenay-Malabry [email protected], [email protected] ABSTRACT: Projects are dealing with bigger stakes and facing an ever-growing complexity. Project risks have then increased in number and criticality. Lists of identified project risks thus need to be decomposed, for smaller clusters are more manageable. Existing techniques are mainly based on a well-known parameter such as the nature of the risk or its criticality. But some limits have appeared since project risk interactions are not properly considered. Project interdependent risks are indeed often considered and managed just as if they were independent. We thus propose an interactions-based clustering method with its associated algorithms and heuristics. Our objective is to group risks, so that the project risk interaction rate is maximal inside clusters and minimal outside. The final objective of this study is to facilitate the coordination of complex projects by reducing interfaces when dealing with risks. We first model project risk interactions through a matrix representation. Then, the overall mathematical formulation of the problem is presented. A case study in the construction industry is finally presented and permits us to propose global recommendations, conclusions and perspectives. KEYWORDS: Project management, Risk, Complexity, Interactions, Clustering. 1 INTRODUCTION A project is a temporary and unique endeavour undertaken to deliver a result, which generally corresponds to the creation of a unique product or service which brings about beneficial change or added value (PMI 2004). A new organization within the firm is then needed to perform a project: new organization new processes which must answer project finalities and objectives in terms of values creation must be set up. These new processes are performed thanks to resources (notably project actors) which belong to the created project organizational system. A project is in essence unique, which means that the project organizational system is to be conceived for each project within a firm (as it is specific to a project). Project organizations are thus in essence temporary organizations. They coexist with permanent organizations which exist within the firm. This coexistence (involving interfaces and dependencies) makes project and project management all the more complex. Moreover, the conception of the project organizational system follows the steps of project phases’ identification and analysis, planning and monitoring. As a consequence when thinking at projects in terms of systems following several phases, many dependencies and interdependencies between phases, sub-systems and other entities can be identified Project systems are indeed in essence complex, be it only through the fact they are performed by project actors, i.e. people (Vidal and Marle 2007; Vidal and Marle 2008). Focusing on the management aspect, it must be kept in mind that management is indeed composed of decisions and activities made by people, those decisions being made at a given instant to reach an objective in the future. Once made, a decision changes the states of the elements it impacts and thus the state of the project itself, targeting a final state for the project (composed of the objectives of the project). The difference between the targeted state and the reached state basically accounts for the project performance. Moreover, every management decision is relative to a context which is the known present situation (resulting from the past decisions). Finally, decisions at a time T help to reach the future objective, which is more or less correctly defined and more or less stable. This overall decisions chaining determines the trajectories of the evolution of the project system. This intrinsic complexity of project management makes it impossible to visualise and manage projects as a whole, notably because of the existence of project complexity induced risks. In order to illustrate this issue of complexity driven risks, let us consider an example, the case of a project within the field of automotive industry. A change in the design of the windscreen (in terms of inclination of this windscreen) implied changes at the connection with the front structure of the car. This was due to the interdependence of the final product components and it provoked in the end several changes which were localised at an interface. But this change also implied rework and a global increase of the duration and cost of some tasks (change of various parameters). This means that the first change had spread throughout the entire system, making it impossible to foresee its evolution properly. As a whole, project management appears to be a complex and risky activity, which underlines the need for MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia efficient and effective project risk management. As a consequence, this paper proposes an innovative method and its associated tools to assist project risk management under complex contexts by focusing on project risk interdependencies. Our goal is to group risks into clusters in order to catch inside of them most of project interactions, which is notably to facilitate the coordination of the project risk management process. 2 CLASSICAL RISK GROUPING METHODOLOGIES IN PROJECTS 2.1 Classifying project risks by nature and/or value Project risk management is classically decomposed into four successive major steps: risk identification, risk analysis, risk response planning and risk monitoring (PMI 2004). Risk identification is the process of determining events which, may they occur, could impact positively or negatively project objectives. Risk identification methods are classified according to two different families: direct or indirect risk identification (Raz and Hillson 2005). This step in the end generates a list of risks. The number of risks in this list may vary from some decades to some hundreds of risks. It is then mandatory to decompose this list into subgroups in order to have more manageable items. During risk analysis, risks are prioritized, essentially according to their probability and impact. At this stage, there are two main types of risk analysis: qualitative and quantitative analysis. The main output of risk analysis is a list or graph, which enables decision-makers to categorize risks as high, medium or low. Risk evaluation scales are often defined in terms of criticality, which is generally a function of probability and impact. Next steps are risk response planning and monitoring. We argue that these steps should be performed after an innovative project risk analysis based on risk interactions since current methods have shown their limits. 2.2 Limits of existing approaches Indeed, the initial goal of clustering processes is to facilitate the coordination and management of risks. Fieldwork proves us this is not always the case with existing methods. Namely, project complexity, described notably in (Baccarini 1996; Edmonds 1999; Laurikkala et al. 2001) involves issues in decision-making under complex situations (Phelan 1995), (Earl et al. 2001). The complexity of a project makes it impossible to visualize simultaneously the complete project (global vision) and all the interactions in the project (Marle 2002). This can notably be underlined when looking at projects through systems thinking (Simon 1981; Le Moigne 1990). Referring to complementary works (Sinha et al. 2001; Aissa 2004), many factors related to project interdependencies have been identified as drivers of project complexity, and thus of risks. But, there are still some phenomena which are not taken into account by classical project risk management methodologies, such as loops or non-linear couplings. Actually, whatever the criteria used for the decomposi- tion of an initial risk list, and whatever the rigour and detail level used, there will always be interactions between risks which do not belong to the same cluster. The problem with current methodologies is that project risk interactions are not clearly included, e.g. in Figure 1, where some links are existing, though not managed (dotted lines). Risks are indeed interrelated with complex links. A previous study we had conducted about 23 risk analysis methodologies enabled to identify complexity-related issues. For instance, there may be propagation from one « upstream » risk to numerous « downstream » risks, the climax of this phenomenon being the famous dangers of the domino effect. Another example may be the existence of loops: amplifying loops are a great danger during projects and are all the more complicated to understand since the nature of the risks which exist within a loop is likely to be different. Figure 1. Classification of projects risks by nature and/or by value Project management current techniques include classical principles underpinning scientific management: the fragmentation of work and the maximization of visibility and accountability. We can argue that today projects are generally managed with single-link trees (WBS, pert, OBS, risk lists) and not as networks (Marle 2002). In the case of risk management, most of the methods use lists, screening or sorting risks, as seen before. Traditional methodologies are mainly single-risk oriented, analyzing their multiple causes and multiple consequences. However, some works have been done to model interdependencies between risks. Bayesian networks for instance link several risks, from multiple inputs to multiple outputs, but they have specific validity conditions: links must be oriented, and there must not be any loop. That means that in some cases, they fail to reflect the real complexity of relationships between project risks. There is thus crucial need for better awareness, consideration and management of project risks, knowing they are intertwined. We propose in this article such a methodology. Our ambition is not to give “exact” results: we want to assist day-do-day project risk management thanks to our method. This one is notably not based on the mathematics of probabilities. It can thus take into account easily the existence of loops and non-linear couplings for instance. 2.3 Overall problem setting and methodology As shown by the former paragraphs, risks are managed thanks to the elaboration of smaller clusters. At this stage, a management issue arises, since decisions may be MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia blocked, slowed down or ineffective if interactions are poorly taken into account. Our research problematic is thus to propose a new additional clustering methodology, which could take into account interactions between risks, in terms of existence and strength. First, we identify possible risk interactions. The whole is synthesized thanks to binary matrix representation. The matrix is then transformed to be a numerical one thanks to the use of the Analytic Hierarchy Process (AHP) principles. These numerical data permit us to develop a linear programming and two approximate iterative algorithms. We then express how these results can still be refined thanks to the introduction of a distance measure and similarity identification process. All the obtained results are then compared to classical decompositions, notably thanks to a case study in the entertainment industry. We then propose some conclusions and call for future perspectives of research around this issue, after studying the implications of our works on day-to-day management. 3 CATCHING PROJECT RISK INTERACTIONS The aim of this part is to build up a binary matrix which can permit to catch and represent project risk interactions and their strength. 3.1 The Risk Structure Matrix (RSM) First, the Risk Structure Matrix (RSM) is a binary matrix is built on the principles of Design Structure Matrix (DSM). The classical partitioning algorithm for grouping elements in this type of matrix is presented. 3.1.1 The Design Structure Matrix (DSM) approach The Design Structure Matrix (DSM) represents and visualizes relations and dependencies among objects. The same objects are both in the rows and columns of the square matrix, which is square. The DSM was introduced by Steward (Steward 1981) with tasks and was initially used basically for planning issues (Eppinger et al. 1994). Since, it has been widely used with other objects, like product components, projects or people (Eppinger and Salminen 2001; Sosa et al. 2004; Danilovic and Browning 2007; Sosa 2008). As for us, we propose to use the concept of DSM for other objects, which are risks, in the context of project management. As tasks, projects and people, project risks are (or can at least be supposed as): • in a finite number (since a project is in essence temporary, with finite resources, objectives, means, etc., i.e. a finite number of elements), • managed during the project management process, • interrelated, (notably because of project and project management complexity factors (Vidal and Marle 2008)) which justifies the use of a methodology for complex interactions management. The reader may note that in this paper, we define risk interaction in terms of the existence of a possible cause and effect relationship between two risks Ri and Rj. We define the binary Risk Structure Matrix (RSM) as the square matrix with RSMij=1 (else 0) when there is an interaction from Rj.to Ri. Main advantages of this approach is to overcome the display issue of complex network and to permit easier calculations which are inherent to the matrix format (eigenvalues, matrices product, matrix transposition, ...). 3.1.2 Building up and partitioning the RSM In order to build the RSM, we need to identify the interactions existing between project risks. The iterative procedure we use is notably addressed in ongoing publications. Classically, the DSM is re-ordered in a way which permits to show first-level blocks, thanks to the wellestablished partitioning process (Steward 1981; Eppinger et al. 1994). This one applied to the RSM gives information about: • the dependent risks: they are engaged in a potential precedence relationship, • the interdependent risks: they are engaged in mutually dependent relation, directly or with a bigger loop, • the independent risks: the risks are basically non-related. The aim of this process is basically to obtain a lowertrigonal by blocks matrix. Partitioning enables to isolate interdependent risks, but our purpose is different. We aim at grouping risks in clusters with maximal internal interactions and minimal inter-clusters interactions. As seen in part 3.3, the partitioning algorithm does not permit to achieve our goal when the structure is very complex. Indeed, it tends to give very few, but big, blocks which are difficult to manage. 3.2 Transforming the RSM into the Risk Numerical Matrix (RNM) Five steps are necessary to carry out our research and build up the Risk Numerical Matrix (RNM) which takes into account the strength of risk interactions. 3.2.1 Step 1: Individual sub-problems The presence of a 1 in the binary RSM expresses the existence of a possible cause and effect relationship between risks Ri and Rj. RSMij=1 implies two possible ways to address the situation: this can be seen either as a possible risk input of Ri coming from Rj, either as a possible risk output from Rj reaching Ri. Similarly as in (Chen and Lin 2003) for design tasks, we combine these visions. Two stages must thus be performed. For each Ri, we isolate the risks which are related with Ri in column (possible effects) and in row (possible causes). They are called the Binary Cause or Effect Vectors and are relative to one risk Ri (BCV|Ri and BEV|Ri). 3.2.2 Step 2: Evaluating the interactions’ strength We buid up two matrices regarding the risk Ri based on the two previously isolated sets of risks (in rows and in columns), which constitute the set of alternatives. They are called Cause or Effect Comparison Matrices and are both relative to one risk Ri (CCM|Ri and ECM|Ri) (see Figure 2). They permit to summarize the strength of risk interactions which are related to the risk Ri which is MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia successively studied as a cause or an effect for other project risks. Fieldwork has proved us that sometimes, people find it easier to say that a cause C1 is more likely to produce an effect E than another cause C2, or similarly, that an effect E1 is more likely to be the consequence of a cause C than another effect E2. Also given that the Analytic Hierarchy Process (AHP) numerous applications in the field of project risk management (Gourc and Bougaret 2000), this led us to claim for the use of the AHP-based principle of pairwise comparisons to assess project risk interactions (as we define them in this article). The Analytic Hierarchy Process (AHP) was developed by Thomas Saaty (Saaty 1977; Saaty 1980 ; Saaty 2000; Saaty 2003). It is a multi-criteria decision-making method. It permits the relative assessment and prioritization of alternatives. The AHP is based on the use of pairwise comparisons, which lead to the elaboration of a ratio scale. The AHP models the problem as a hierarchy, consisting of an overall goal, a group of alternatives, and a group of criteria which link the alternatives to the goal. Pairwise comparisons are carried out by asking how more valuable an alternative A is to criterion C than another alternative B. Pairwise comparisons finally constitute square matrices, the values of which are between 1/9 and 9, and the diagonal elements of which are equal to 1 while the other elements verify two conditions: • The i-jth element is equal to the comparison between i and j regarding the considered criterion. • For i different from j, the i-jth element is equal to the inverse of the (j-i) element. This piece of information is processed mathematically, in order to transform user information into mathematical one. Priorities are then determined thanks to these matrices and a global consistency test can be performed to evaluate the coherence of the judgements. The final result is a table which gives a global evaluation of each alternative for the objective and for each criterion. In our case, we have two parallel pairwise comparison processes to run. The first one consists in the ranking in rows for each project risk. Given the risk Rk, the set of alternatives are all the non-zero elements of risks other than the diagonal element in row k. The criterion on which the alternatives are evaluated is the contribution to Rk in terms of risk input: in other terms, for every pair of risks which are compared, Ri and Rj (thus following RSMki=RSMkj=1), the user should assess which one is more important to risk Rk in terms of probability to be a risk input (i.e., a cause) for risk Rk. Numerical values express these assessments thanks to the use of the traditional AHP scales. The second one is the ranking in columns, according to the same principles. 3.2.3 Step 3: Consolidating the results When performing the pairwise comparison-based evaluation, eigenvectors of each matrix ECM|Ri and CCM|Ri are then to be calculated. It enables to find the principal eigenvectors which are related to the maximal eigenvalue. They are called the Numerical Cause or Effect Vectors, as shown in figure 2 and are relative to one risk Ri.(NCVi and NEVi). Consistency of the results should be tested. 3.2.4 Step 4: Aggregating the results For each risk Ri, Numerical Cause or Effect vectors are respectively aggregated into Numerical Cause / Effect Matrices (NCM and NEM, see figure 2). The i-th row of NEM corresponds to the eigenvector of CCM|Ri, which is associated to its maximum eigenvalue. The j-th column of NCM corresponds to the eigenvector of ECM|Rj, which is associated to its maximum eigenvalue. 3.2.5 Step 5: Compiling the results The two previous matrices are aggregated into a single Risk Numerical Matrix (RNM), the values of which assess the relative strength of local interactions (figure 2). The RNM is defined by a geometrical weighting operation (based on the possible assumption that both estimations can be considered as equivalent). RNM (i, j ) = NCM (i, j ) × NEM (i, j ) (1) The RNM thus permits to synthesize the existence and strength of local precedence relationships between risks. The reader should note that, for all practical purposes, another possibility to evaluate the strength of these interactions is to perform a direct evaluation on 10 level Likert scales. This evaluation is less time and resource consuming but it requires having greater expertise and a very global vision of the project. 4 PROPOSED CLUSTERING METHODOLOGY AND REFINEMENTS 4.1 Problem definition We want to cluster risks to maximize intra-cluster interactions thanks to the use of the RNM (Hartigan 1975; Fowlkes and Mallows 1983; Murtagh 1983). We do insist on the fact that the values of the RNM are local judgements, which implies that risk interactions assessments are in essence relative. However, this first clustering is useful, since it permits to focus on the most significant local risk interactions. A set of risks is in essence a complex one, since interactions do exist between them. Let us suppose we know the RNM of this set of risks (the former steps to build the RNM should have been followed by the user). Let K be the number of clusters of the optimal clustering solution, which maximises intra-cluster global interactions value. This INTRA value is defined by the sum of the values of all interactions between risks which belong to a same cluster. The INTER (Inter-cluster global interactions) value is defined by the sum of the values of all interactions between risks which are not paired inside a same cluster. The sum of INTRA and INTER values corresponds to the sum of all risk interactions values, which is constant. As a consequence, maximizing INTRA is equivalent to minimizing INTER. We do not know K in advance. However, we know some constraints about K. MOSIM SIM’10 - May 10-12, 2010 - Hammamet - Tunisia Figuree 22. How to build the Risk Numerical Matrix (RNM) K Namely, the goal is to assign project members me to clusters in order to manage more properly the risks ris which belong to a same cluster, i.e. which are strongly gly interdependent. It is known that people have a limitedd capacity to manage simultaneously numerous objects. s. W We introduce the variable Smax as the maximum size off a cluster. We indeed choose to leave some margin comp mpared to the classical empirical rule of 7 objects to bee managed m simultaneously. This permits us to know a lower low bound of K, which is N −1 K min = INT ( ) +1 S max (2) where INT is the integer part of a reall nu number. ming problem 4.2 Formulating the linear programm Here is the corresponding integer progr gramming problem formulation. This problem is to be solve lved for each value of K which is superior to Kmin. We first fi introduce the following decision variables: ∀i,1 ≤ i ≤ N , ∀k ,1 ≤ k ≤ K , xik = 1 (3) if risk Ri belongs to cluster Ck. The objective function, which is to bee maximized, is as following in Eq. (4) (which is not linear ar for the moment) K N k =1 i =1 j =1 Eq. 5 and 6). Problem constraints are the following (E (4) (5) k =1 as we argue for clusters disju sjunction in order to permit easier management in practice. ce. N ∀k ,1 ≤ k ≤ K , ∑ xik ≤ S max m (6) i =1 In our case, we fix Smax=9, in order to permit the future management of the clusters wh when cluster owners will be designated. This problem is not linear bu but we can make it easily linear thanks to the introducti ction of new decision variables (Eq. (7)) and new constra traints (Eq. (8)). ∀i, ∀j ,1 ≤ i, j ≤ N , ∀k ,1 ≤ k ≤ K , yijk is binary (7) We finally define yijk by adding ing the constraints: ∀i,1≤ i ≤ N,∀j,1≤ i ≤ N,∀k,1≤ k ≤ K, yijk ≤ xik + xjk −1 (8) This equation forces yijk to bee eequal to 0 if xik and xjk are not both equal to 1, i.e. if Ri and an Rj do not belong to the same cluster. All other constra straints are kept for problem formulation. The objective ffunction can then be rewritten thanks to these new decision dec variables, in Eq. (9). K N INTRA = ∑∑∑ xik x jk RNM (i, j) j ∀i,1 ≤ i ≤ N , ∑ xik = 1 N N INTRA = ∑∑∑ yijk RNM NM (i, j ) (9) k =1 i =1 j =1 We use OPL (Optimization Programming Pr Language) to solve this problem. However, r, its complexity is high (2N1 ), and problems over 20-211 rrisks appear to be critical when testing them. That is why hy we have based our work MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia on some easy heuristics, which permit us to approximate the optimal solution of the problem. 4.3 Using heuristics to answer this problem Both of these approximate algorithms are iterative, but they use two different values for clustering conditions, as described in Eq. (10) and Eq. (11). The first iterative algorithm IA1 is based on the maximum value between separate clusters. The second one IA2 is based on global interactions value between clusters. In the two cases, these values are to be maximized at each step. Val1 (Cα , Cβ ) = Val2 (Cα , Cβ ) = max RNM (i, j ) (10) i∈Cα , j∈C β ∑ RNM (i, j) + RNM ( j, i) (11) i∈Cα , j∈Cβ At the initial step, all risks are isolated: every initial cluster is a singleton. The maximum value is obtained for two isolated risks Ri0 and Rj0, which are grouped into a first cluster C1. At each following step, the previous value (Value1 or Value2) is maximized. This procedure is repeated iteratively until reaching a solution which respects all the constraints. In the case the maximum size of a cluster is reached before the end of this procedure, the second maximum value in the RNM is identified and the clustering operation is done on the corresponding interaction. The performance of these heuristics was notably studied in other publications (Vidal and al., 2009). 5 CASE STUDY 5.1 Introduction The following case is a large infrastructure project, which consists in the implementation by a French company of a new tramway transportation system in a city. The goal is to build both the infrastructure and the transportation system. The customer is a 750 000 inhabitants city in a country C. This notably comprises: • The construction of a depot to stock trains and execute their control and maintenance • The installation of tracks throughout the city, with issues about altitudes and high slopes, • The construction and onsite delivery of the corresponding trains. • The establishment of a traffic signalling system, interconnected to the existing system dedicated to road. An industrial partner realises the civil work which is to permit the installation of the tramway. The project was initialized by the government of country C in 1995. The first selections of the firms which would execute the project occurred in 1999. The project contract was signed in 2002. After negotiations with banks, the government and the future operator (in which the French firm which executes the project holds shares), the final concession contract was signed in 2004. The project started in February 2005, with a practical start of the execution in 2006. Until now, a project risk management process has been carried out and led to the existence of 8 lists of risks which nurtured the successive risk reviews. We focus here on the System product line, which considers the integration of all the aspects of the project, and is thus to be one of the most complex ones. The corresponding risk list (42 risks) we have been working on can be seen afterwards on Figure 3. The 42 risks which are present in the list are very diverse and are classified according to six risk classes (risk nature). Risk ownership in terms of responsibility is dispatched to 12 actors in the project. Currently, risk management presently receives moderate attention within the firm and the following issues are to be underlined. First, risk lists are elaborated since they are to be done, but no real attention is paid to them and they are not used as much as they could be. Risk management is still too often considered as an academic work which in not necessary for day-to-day project management. Moreover, risk owners (in terms of responsibility) may sometimes be defined too quickly, since the examination of this list underlines that some ownerships should be rearranged. Indeed, risk owners belong to varied hierarchical levels in the firm structure, and some risk owners are responsible for one risk while other ones are responsible for more than ten. 5.2 First results Our work through the consideration of risk interactions has globally created more inclination and confidence with the use of risk management approaches. The first remark is that when performing the study thanks to the iterative process of risk interaction identification, new risks appeared. Namely, they were consequences or causes of some already identified risks, or they were seen as intermediary risks to explain the link between two existing risks. As a whole, 13 risks were newly identified for a lack of their presence in the list appeared (see Figure 4), which represents an increase of nearly 31% in the number of identified and considered risks. Finally, 6 of the risks which were present in the initial list (R1, R8, R11, R15, R23, R34) were considered as poorly defined or possibly negligible. MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 Safety studies Liquidated damages on intermadiate milestone and delay of Progress Payment Threshold vehicle storage due to depot delay Vandalism on site Traction/braking function : behaviour in degraded mode on slope Local laws and regulations Traffic signalling, priority at intersections Unclear Interface with the Client, for Infra eqt Delays due to client late decisions Travel Time performance Limited Force majeure definition Operating certificate Reliability & availability targets Permits & authorisations Insurance deductibles Archeological findings Discrepancies Client / Operator / Concessionaire CW delay & continuity Responsibility of client on CW delay On board CCTV scope Noise & vibration attenuation Potential risks of claim from CW partner Harmonics level Non compliance contractual Rolling Stock Non compliance technical specs Rolling Stock Exchange risk on suppliers Track installation equipment performance Tax risk on onshore more poles Security requirements Track insulation Delay for energising Fare collection requirements Construction safety interfaces Electromagnetic interferences Exchange risk Risk of partial rejection of our request for EOT Interface rail / wheel Risk on Certification of our equipement OCS installation Banks stop financing the project Costs of modifications not covered by EOT agreement SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS Actor A Actor B Actor A Actor C Actor A Actor A Actor D Actor E Actor E Actor D Actor B Actor B Actor D Actor B Actor F Actor B Actor G Actor H Actor B Actor I Actor D Actor B Actor D Actor A Actor A Actor F Actor J Actor F Actor D Actor E Actor K Actor D Actor G Actor C Actor E Actor F Actor B Actor E Actor L Actor C Actor B 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 CLASS 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Technical Contractual Contractual Contractual Technical Contractual Contractual Contractual Contractual Technical Contractual Contractual Technical Contractual Financial Contractual Contractual Contractual Contractual Technical Technical Contractual Technical Technical Contractual Financial Client/Partner/Sub-contractor Financial Contractual Technical Technical Project Management, Construction site Contractual Technical Technical Financial Contractual Technical Country Project Management, Construction site Contractual Contractual COMM ON RISK OWNER IN D IV ID U A L PRODUCT LINE D E L IV E R Y (in m o n th s) RISK IDENTIFICATION IMPACTED RS8 Q U A L IT Y - S C O P E Review Session Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Common Individual Individual Individual Individual Figure 3. Initial risk list of the System product line of the project NEW RISKS The identification of the existing risk interactions was thus performed. A direct evaluation on a 10 level Likert scale of the strength of interactions was executed, although we recommend the use of AHP in most cases, due to the high expertise of interviewees. An observation is that there were some difficulties while performing this step since: • This step is to require the participation of several experts of the project for it implies a very wide view of the project elements and stakes. • Some bias may be included in the evaluation of interactions since it appears that interactions are often thought at a first sight in terms of impact and not in terms of precedence. Great attention should thus be paid to that point in order to analyse the results. Return profit Extra trains Pedestrian zones Train performance Waiting time at stations Depot delay Survey Ticketing design delays Track installation Reengineering / Redesign Slabs pouring Initial specifications of CW Available cash flow N° 43 44 45 46 47 48 49 50 51 52 53 54 55 Figure 4. Newly identified risk thanks to the risk interaction identification process MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia In the end, a global Risk Numerical Matrix for the studied risk network was obtained and analyzed in next part. 5.3 Deeper analysis and discussion Even when separating in its connected components, the LP-problem was too large to be solved by OPL. The use of heuristics was thus necessary. The clustering iterative algorithm IA2 was performed to obtain a first good approximate result for the clustering operation. Indeed, algorithm IA2 proved to offer better results in other cases and that is why we chose it in that case (Vidal et al. 2009). As a whole, the following clusters were obtained (Figures 5 and 6). Some risks appear to be high accumulation risks, notably the budget related ones in terms of return profit (R43) or risk of rejection of extension of time EOT (R37) and liquidates damages (R2). These ones are to be considerably monitored since many paths in the risk network are likely to lead to them. Same observation can also be made for travel time performance (R10). The obtained clusters seem to be quite consistent with the fieldwork as they form groups of risks which seem to be relevant in order to assist project risk management. Cluster C3 and C4 for instance permit to group possible chain reactions which could imply delay (respectively for the permits and authorizations, and for the depot construction and track installation). This appears to be all the more interesting than such chain reactions were not highlighted and managed before during the project, essentially because they mix schedule, technical, budget and human aspects. For instance, there were no discussions between Actor A and Actor E regarding the link between R3 (Vehicle storage due to depot delay) and R32 (Delay for energising), whereas this interface should have been particularly highlighted retrospectively. One commentary is that cluster C1 should however perhaps be separated into two parts by regrouping all financial risks in a sub-cluster. This appears all the more relevant than these financial risks are also linked to many other risks which exist in other clusters. Therefore, managing them as a complete cluster could for sure be very interesting. Another issue which arises is the question of risk ownerships. Indeed, it appears that within clusters, there are numerous risk owners, and often numerous risk classes. One question which is to be addressed is how coordination can be facilitated since there seems to be some benefit to discuss with all the risk owners within a same cluster. One thing which was suggested is that a meeting with all the impacted risk owners of a cluster could permit to nominate / vote for a responsible for the cluster who could facilitate the coordination between the interrelated risks. One of the possible nominees for this cluster responsibility could be the least common boss in the hierarchical structure of the project. Moreover, new constraints might be added to perform more clustering solutions. For instance, new tests are to be conducted by varying the maximum possible size of a cluster. Another constraint which could be added would also be to add a maximum number of different risk owners within a cluster. When comparing with classical approaches, clustering by interactions permits an important improvement re- garding the consideration of interactions. Indeed, the intra-cluster value is increased by as much as 61% when comparing with classical clusterings. As a whole, the feedback with this case study is that in order to obtain helpful results thanks to this methodology would be in the end to: • Perform pertinent risk identification and risk interactions identification and evaluation processes (in group) in order to obtain a good description of the situation • To have a same hierarchical level in the risk structure to study same level risks in the chain reactions. • Identify carefully during the initial step the correct risk owners, i.e. the actors which seems initially the most appropriate ones to hold the responsibility for each risk. • Perform the clustering operation thanks to the iterative algorithms or with OPL on the LP problem if processing time can be improved thanks to some operations. • Analyse the obtained results and identify possible chain reactions, possible accumulation risks and the actors which are to be responsible for each risk cluster in order to facilitate the global coordination of the project risk management process. 6 CONCLUSION As a whole, this article presents innovative tools based on the integration of risk interactions in the processes of risk analysis and risk clustering for efficient project risk management. As shown by the case studies, the tools which are proposed here permit greater communication on project risks as well as a better confidence in risk management activities thanks to two aspects at least. • First, the evaluation of risk interactions which is performed when building up the RNM implies a two-step process (looking in terms of causes, and then of consequences). Information can thus be checked and refined since one interaction should be listed twice (from cause to effect, and from effect to cause): this checking process permits a better confidence in risk identification and risk interaction identification. • Moreover, clustering risks in order to maximize intra-cluster global interactions value permits to facilitate the coordination of risk monitoring and controlling activities, as it underlines the need for cooperation and transversal communication within the project team. It permits greater communication between people, since it does not seek the identification ownership, responsibility and/or accountability, but the identification of risk interdependencies. After the clustering process, coordination is made by the person who is assigned to the cluster, but communication has been facilitated before, meaning we have less defensive phenomena. MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia RO Initial Actor B Actor D Actor J Actor B Actor B Class Initial Contractual Technical Client/Partner/Subcontractor Contractual Contractual C1 Liquidated damages on intermadiate milestone and delay of Progress Payment Threshold Travel Time performance Appitrack performance Risk of partial rejection of our request for EOT Banks stop financing the project Return profit Extra trains Waiting time at stations Available cash flow RO Initial Actor B Class Initial Contractual C2 Potential risks of claim from CW partner Initial specifications of CW RO Initial Actor E Actor B Actor G Actor G Class Initial Contractual Contractual Contractual Contractual C3 Delays due to client late decisions Permits & authorisations Discrepancies Client / Operator / Concessionaire Fare collection requirements Ticketing design delays RO Initial Actor A Actor B Actor H Actor B Actor D Actor E Class Initial Contractual Contractual Contractual Contractual Technical Project management, Construction site C4 vehicle storage due to depot delay Archeological findings CW delay & continuity Responsibility of client on CW delay Noise & vibration attenuation Delay for energising Depot delay Track installation Slabs pouring RO Initial Actor A Actor D Actor B Actor D Actor E Actor D Actor L Class Initial Contractual Contractual Contractual Technical Contractual Technical Country C5 New local laws and regulations Traffic signalling, priority at intersections Operating certificate Reliability & availability targets more poles Security requirements Risk on Certification of our equipement Reengineering / Redesign RO Initial Actor A Class Initial Technical C6 Traction/braking function : behaviour in degraded mode on slope Train performance Survey RO Initial Actor I Actor C Class Initial Technical Project management, Construction site C7 On board CCTV scope OCS installation RO Initial Actor A Actor D Class Initial Contractual Technical C8 Non compliance technical specs Rolling Stock Interface rail / wheel RO Initial Actor A Actor C Actor E Actor B Actor F Actor E Actor A Actor F Actor F Actor K Actor C Actor F ??? Class Initial Technical Contractual Contractual Contractual Financial Technical Technical Financial Financial Technical Technical Financial Contractual Isolated risks Safety studies vehicle storage in Bellevue due to depot delay Unclear Interface with the Client, for Infra eqt Limited Force majeure definition Insurance deductibles Harmonics level Non compliance contractual Rolling Stock Exchange risk on suppliers Tax risk on onshore Track insulation Construction safety interfaces Exchange risk Costs of modifications not covered by EOT agreement Pedestrian zones Figure 6. Results of the clustering operation Figure 7. The clustered project risk network MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia However, this implies that a shift should be operated in the skills of project risk managers (or at least the team members who are in charge of the management of the obtained clusters). Such project team members should indeed be able to facilitate communication and to show great adaptability since they need to manage risks which are to be of different nature. As a whole, this article permits to make a comparison between several possibilities for grouping risks in a project. Our aim is not to criticize the use of classical approaches: on the contrary, we refer to them as points of comparison and claim for the use of conjoint classifications which can all give powerful insights on reality. Lots of aspects of this work and its results may however be discussed. We thus identify several research perspectives to consolidate this approach. • Challenging the definition of risk interaction and trying to integrate other risk characteristics (than probabilities and precedence relationship) into the definition of risk interactions. • Evaluating with more reliability the relative weights of risks. The sensitivity of this evaluation should first be explored. • Exploring more sophisticated graph partitioning heuristics and multi-objective clustering. • Exploring new constraints to perform other clustering operations. For instance, the maximal size of the clusters may vary. One could think of asking for a density constraint or asking that the obtained clusters may as a whole be of a similar size. Finally, new case studies are to be performed in order to validate even more this approach and study both the practical applications (and improvements) thanks to these results and the future implications on project management processes and organization. REFERENCES Aissa, A. (2004). Formalisation et quantification de processus de gestion des interactions des projets. Laboratoire Genie Industriel, Ecole Centrale Paris. Baccarini, D. (1996). "The concept of project complexity – a review." International Journal of Project Management 14(4): 201-204. Chen, S.-J. and L. Lin (2003). "Decomposition of interdependent task group for concurrent engineering." Computers and industrial engineering. Danilovic, M. and T. Browning (2007). "Managing complex product development projects with design structure matrices and domain mapping matrices." International Journal of Project Management 25: 300-314. Earl, C., J. Johnson, et al. (2001). Complexity in planning design processes. 13th International Conference on Engineering Design, Glasgow, Scotland. Edmonds, B. (1999). Syntactic measures of complexity. faculty of arts. Manchester, University of Manchester. PhD. Eppinger, S. and V. Salminen (2001). Patterns of product development interactions. International Conference on Engineering Design, Glasgow, Scotland. Eppinger, S., D. E. Whitney, et al. (1994). "A modelbased method for organizing tasks in product development." Research in Engineering Design(6): 1-13. Fowlkes, E. and C. Mallows (1983). "A method for comparing two hierarchical clusterings." Journal of the American Statistical Association 78(383): 553-584 Gourc, D. and S. Bougaret (2000). "Le management des risques appliqué au management des projets et des portefeuilles de projets en recherche et développement." Revue Travail et Méthodes n° 552: pp 35-42. Hartigan, J. A. (1975). Clustering Algorithms. New York, John Wiley & Sons. Laurikkala, H., E. Puustiner, et al. (2001). Reducing complexity of modelling in large delivery projects. International Conference on Engineering Design, Glasgow, Professional Engineering Publishing. Le Moigne, J.-L. (1990). La théorie du système général. Théorie de la modélisation, Presses Universitaires de France. Marle, F. (2002). Methods for helping decision-making in projects. PMI Europe Conference, Cannes, France. Marle, F. (2002). Modèle d’informations et méthodes pour aider à la prise de décision en management de projets. Génie Industriel. Paris, Ecole Centrale Paris. PhD: 220. Murtagh, F. (1983). "A survey on recent advances in hierarchical clustering algorithms." The Computer Journal 26(4): 354-359. Phelan, S. (1995). From chaos to complexity in strategic planning. 55th meeting of Academy of Management, Vancouver PMI, S. C. (2004). A guide to the project management body of knowledge (PMBOK) (2004 ed.). Newton Square, PA, USA. , Project Management Institute. Raz, T. and D. Hillson (2005). "A comparative review of risk management standards." Risk Management: An international journal 7 (4): 53-66. Saaty, T. (1977). "A scaling method for priorities in hierarchical structures." Journal of Mathematical Psychology 15: 234-281. Saaty, T. (1980 ). The Analytic Hierarchy Process: Planning, Priority Setting, Resource Allocation, McGraw-Hill. Saaty, T. (2000). Fundamentals of the Analytic Hierarchy Process. Pittsburgh, PA, RWS Publications. MOSIM’10 - May 10-12, 2010 - Hammamet - Tunisia Saaty, T. (2003). "Decision-making with the AHP: Why is the principal eigenvector necessary." European Journal of Operational Research 145(1): 85-91. Simon, H. (1981). The Sciences of the artificial. Cambridge The MIT Press. Sinha, S., A. I. Thomson, et al. (2001). A complexity index for the design process. International Conference on Engineering Design, Glasgow, Professional Engineering Publishing, Bury St Edmunds. Sosa, M. (2008). "A structured approach to predicting and managing technical interactions in software development." Research in Engineering Design. Sosa, M., S. Eppinger, et al. (2004). "The Misalignment of product architecture and organizational structure in complex product development." Management Science 50(12): 1674-1689. Steward, D. (1981). "The Design Structure Matrix: a method for managing the design of complex systems." IEEE Transactions in Engineering Management 28(3): 71-74. Vidal, L. and F. Marle (2007). Modeling project complexity. International Conference on Engineering Design, Paris, FRANCE. Vidal, L. and F. Marle (2008). "Understanding project complexity: implications on project management." Kybernetes, the International Journal of Systems, Cybernetics and Management Science. Vidal, L., F. Marle, et al. (2009). Interactions-based clustering to assist project risk management. International Conference on Engineering Design ICED'09, Stanford, USA.