The case study for this essay comes from the Institute for Operational Research and Management Science (INFORMS) Interfaces journal, in the article ‘Dispatch Optimization in Bulk Tanker Transport Operations’ (Gifford et al. 2018).

A significant percentage of freight in North America is transported by large-scale for-hire full-truckload carriers. In this situation, carriers are tendered freight orders which originate and are delivered across a wide geographical range of customer locations. There are different service and cost considerations to take into account such as that full-truckload carriers operate on random one-way networks which contrasts with less-than-truckload (LTL) and small package operators (such as UPS, FedEx, etc.). This case study addresses a specific problem in the bulk transport (fuels and chemicals) division of Schneider National Inc.. Schneider National Inc. accept 350 customer orders per day which involves 10,000 distinct commodities. A corresponding freight for this usually comprises of 1,000 drivers and 1,600 tanker trailers. Chemical interaction properties of these commodities impose rather complex product-sequencing constraints such as interorder tanker wash and preparation processes, and the selection of specific trailer configurations. Schneider National Inc. also have to consider the more common fleet dispatch problems in addition to those of the chemical interactions. To address this problem, Schneider National Inc. designed and implemented a multiphase, multidimensional matching algorithm and developed new business processes that enabled business partners to leverage optimised solution recommendations. Since the new system has been implemented, it has delivered a wide range of improvements. A significant improvement in productivity and customer service was seen, as well as a documented amount of over $4 million in annualised operational and capital cost savings.

The main problem addressed in this case study is a dispatch problem, which can be seen as a multidimensional matching problem. The solution technique comprises of interleaved candidate-generation and optimisation processes (Figure 1). This is done in two phases. Phase one comprises of three parts: tanker trailer feasibility, tanker route generation and tanker route optimisation. Phase two comprises of driver work-assignment schedule generation. The objective of the solution system is to make recommendations which are then accepted at a rate of 80-90% but with little to no human interaction involved.

Figure 1: The solution technique comprises interleaved generation and optimization processes.

This is what can be described as a ‘heterogeneous vehicle routing problem’. Many studies in this focus mainly on the development of heuristic approaches and the determination of adequate upper and lower bounds on the optimal solution (Desrochers and Verhoog 1991, Yaman 2006, Li et al. 2007, Euchi and Chabchoub 2010).

A very early definition of a model by Ackoff and Sasieni (1968) states that: ‘a model is a representation of reality’. This is a very simple and succinct way of describing a model, though it does not fully explain what a model is. For a full definition of a model, this simple definition needs to be expanded, whilst accounting for the purpose of the building of the model, that no model can ever be complete and that it is possible to build multiple models for the same, single apparent reality. OR/MS models are also often built to enable a manager to exercise better control or help people to understand a complicated situation, but one must also take into account that an OR/MS model could be of value to people who are not in positions of high power within a business or company (Ritchie et al. 1994). Models are usually modified by an individuals’ experience, sometimes due to the failings of the model and sometimes because they are criticised or challenged by others – the models in OR/MS are more concerned with models which are explicit and external. Taking all of these views on a model into consideration, the more explicit and accurate definition of a model as said by Pidd (2003) is: ‘a model is an external and explicit representation of part of reality as seen by the people who wish to use that model to understand, to change, to manage and to control that part of reality’.

From Pidd’s (2003) definition, we can infer that there are three main aspects to a model. The first aspect is that a model represents or describes something real. The second aspect is that a model simplifies that real entity. The final aspect is that the production of a model has a purpose, generally to make some sort of calculations or predict how the entity will behave (Williams, 2008).

The first two aspects of a model can be otherwise worded as taking a problem from the ‘real world’, trying to simplify this problem and find a way to represent it in a simpler, more comprehensible way. But, we also need to take into account the purpose of the model, which falls under the third and final aspect. A model needs to be built in such a way that it can be manipulated and tell us something useful.

There are two main uses OR/MS models have: the exploration of possible consequences of actions before they are implemented, and to be used as part of embedded computer systems for routine decision support (Pidd, 1999). The model in the aforementioned case study (Gifford et al. 2018) falls under the second use, as it mainly consists of an embedded computer system built for client use.

When building a model, we need to ask seven key questions: why the model is being built, for whom the model is being built, how much of the whole system is being modelled, how general will the model be, how will the causality be modelled, how will the behaviour be modelled and how complex or simple the model should be (Williams, 2008).

A ‘good’ OR/MS model should be transparent, that is, simple to understand, should be accurate and reliable, cost and resource effective as well as reusable and replicable. A ‘good’ model should also have different objectives and clear communication flowing between the stakeholders, the client and analysts working on it.

As Williams (2008) states, there are four main areas of issue when looking at what makes a ‘good’ OR/MS model: understanding the client, relationship with the client, what is being analysed and how the analysis is done. When trying to understand your client, you need to understand the individuals and their bounded rationality, as well as groups, their dynamics and any social effects that might be in place within and between such groups. A group’s shared views and beliefs will influence and be influenced by the strategic decision process of strategy formulation and implementation (Schwarz, 2003). From this stems your understanding of how the manager will make their decision and how they work – whether the manager makes the decision independently or through a team of individuals. If the managers work long hours at a high pace, they are less likely to take notice of the modelling process. “The key to communication is understanding the impatience of the client” (Abdel-Malek et al. 1999).

Gass (1987) mentions the five major criteria for model evaluation (i.e., whether the model is ‘good’ or not) the General Accounting Office (GAO) approach focuses on. These include: documentation, validity, computer model verification, maintainability and usability. The documentation of a model must specify what, why and how it has been done. In other words, the documentation must be detailed enough to allow for replication by independent evaluators to check the validity of the model’s results and claims. To check the model’s validity, there is no set, standard way. In most cases, to check the model validity, any evaluation should encompass the issues and concerns surrounding the theory, data and operations of the model. Following on from the validation of the model, the computer-based model must be verified to check it runs as intended i.e., that the computer program accurately describes the model (as the model was designed to do). If the model is built for long-term use, as many government models are, it calls for a pre-planned and regularly scheduled program to review the model’s accuracy over its life cycle. Such a program includes a process for updating and changing model parameters as well as the model structure. The final major criteria when evaluating a model is checking its usability. This depends on the availability of data, a user’s ability to understand the model’s output, the model’s portability (i.e., that the model can be transferred from one machine or system to another), the associated costs required to run the model as well as required resources.

Using the model evaluation criteria as mentioned by Gass (1987), model specifications mentioned by Williams (2008) and Pidd (2003), and my own observations, we will evaluate the model in the case study (Gifford et al. 2018) mentioned in the first part of this essay. From the full definition of a model, we can infer that a model is a simplified representation of a ‘real world’ problem. The aforementioned case study notes that “the dispatch optimization system is a decision support tool, not a decision-making tool” for which it is “impossible … to accurately capture and adjust to all relevant real-world information”. From this passage, we can presume that though the model specified represents the problem in the ‘real world’ sufficiently, it does not present every possible problem with a relevant solution whilst trying to minimise human interaction in the process completely – one of its objectives is to produce recommendations with “little or no human review” which are accepted at a rate of 80-90%. As previously stated, a model is a simplification of the ‘real world’ problem, therefore if the model produced in this case study was to encompass every possible problem to occur and produce a viable solution to each of these, it would become over complicated – due to over fitting – as well as computationally and resourcefully expensive.

As mentioned earlier, the model specified in this case study is used as a part of an embedded computer system built for the clients’ use. Due to this, a computer based verification stage must be implemented in its evaluation as well as its maintainability and the usability of its user-interface elements. This was done by providing additional elements to the user-interface to allow network managers to adjust parameters and other business-rule conditions that might be relevant to their order. Here, we can also see that there is clear communication between the analysts and programmers of the system and the managers that use the system as it is taking into account the managers’ involvement in both the model building process as well as the implementation of the model. The clear flow of communication between the stakeholders can also be seen as there was a flow of timely and actionable feedback kept throughout the whole process, allowing for frontline planners and dispatchers and their direct managers to evaluate the different phases of the model and modelling process as it was being developed, and allowed them to give their input into the process which could potentially improve the model and make it more understandable by the user’s for whom it is being developed.

There are many different stakeholders involved in the modelling process within this case study: Schneider National Inc. – for whom the model was initially built for, – area planning managers (APMs), computer system programmers and analysts, bulk tanker drivers, office staff, and management personnel (within the company creating the model as well as within the clients’). All of these influence or are influenced by the modelling process and the final model outcome. The model captured the interest of all of these stakeholders (most importantly the management personnel and APMs) as the model would provide three main areas of benefit: cost avoidance, additional revenue opportunity and improvement in the productivity of office staff. The APMs have full visibility to all dispatch activity in the market, based on the output of the computer system built upon the model, which falls under the usability criterion of model evaluation.

The case study goes on to talk about planned improvements to the model implemented. The model implemented was not complete, but as we read back to Pidd’s (2003) definition of a model, we can infer from it that a model is never complete, and therefore adaptable and open to any future improvements and development of additional capabilities, as the model in this case study states.

Though the overall structure, documentation and usability of the model provided is ‘good’, there are a few things which the case study does not mention in great detail, or in some cases at all. Looking back to the key questions to ask during the modelling process (Williams, 2008), we can see that the questions of why the model is being built, for whom, how much of the whole system is being modelled, modelling of the causality and the model’s complexity are being answered, there is little mention of how behaviour of human interaction within the model is being modelled – whether modelling the behaviour of human participation within the model is simply too complex and tedious, or whether this variable is being ignored altogether. One way the case study mentions this variable is that the model is a computer system, therefore any human interaction will be done with the computer system – though this interaction is mainly the clients’ and does not mention any further interaction outside of the use of the computer system even though the objective of the model was to reduce human review whilst increasing the acceptance of recommendations at a rate of 80-90%, not fully taking out human review from the equation.

The project within this case study had many positive outcomes and successes. One such success of the project was that it made the users of the model embedded computer system appreciate the strengths of using a computer-based decision making support system and that, due to the automation of the process, they are able to focus their energy on other projects which have a higher need of expert human intervention for the completion and development.  Within the first few months of the implementation of the solution system, an increase in productivity for assets, drivers, office staff, and management personnel was observed. Another side effect of using the automated system showed significant direct cost savings in fuel and other related expenses. This lead to developing important insights and new ideas about the process of modelling and creating data-driven decision support models. One such insight was the implementation of a feedback mechanism being built directly into the model and computer system rather than being left as an afterthought where it would make it more difficult to describe what the users found to be hard to understand within the system without properly noting their every step throughout the process. At first, this process was met with scepticism from business experts who insisted their work could not be adequately or even modestly automated due to its complexity.

Tilanus (1985) mentions different ways in which the failures and successes of a model can be grouped. These are: client-oriented, OR/MS-oriented and relation-oriented. Each of these groups have their own subcategories (Table 1 and Table 2).

Table 1: Reasons for model/modelling process failure.  Table 2: Reasons for model/modelling process success.

From a technical perspective, the model described in the case study managed to reduce the complexity of the problem presented by Schneider National Inc., which can be seen as a reason for success as shown in Table 2, as an OR/MS-oriented reason. It is also an innovative way to address and solve a practical problem successfully whilst taking into account its complexity and significant size. The model itself is stated to meet the requirements of a ‘real-time dispatch system’ whilst exceeding its business expectations and meeting its theoretical criteria. All of these can be seen as reasons for success as it meets with the description of what a ‘good’ model is and fits in with the evaluation criteria mentioned by Gass (1987).

Before the solution system was put into practice, the process was done mostly manually and was overly reliant on human judgement and expertise, tribal knowledge, and disparate, difficult-to-maintain data sources. With the new system, a significant increase in automation was seen with the use of mathematical optimisation techniques to choose the optimal solution based on a customer’s needs and requirements. This generated significant gains for the company in productivity as well as direct cost savings in fuel and other expenses. These savings were quantified and it was found that the annualised benefits appeared to be: $2.3 million in cost avoidance (empty mile improvement), $2.5 million additional revenue (improved driver productivity) and 28% planner productivity. All three of these benefits exceeded their projected values by 19%, 20% and 18% respectively.

A verification letter is attached to the end of the case study (Gifford et al. 2018) written by John Bozec, Senior Vice President/General Manager of Schneider Bulk Division as a ‘written verification of success in practice’ of the model produced and implemented for Schneider National Inc..

Essay word count: 2,790


  • Abdel-Malek, L., Wolf, C., Johnson, F., Spencer, T. (1999) ‘OR Practice: Survey Results and Reflections of Practising INFORMS Members’, Journal of the Operational Research Society, vol. 50, pp. 994-1003.
  • Ackoff, R. L., Sasieni, M. W. (1968) ‘Fundamentals of Operations Research’, John Wiley & Sons, New York.
  • Desrochers, M., Verhoog, T. (1991) ‘A new heuristic for the fleet size and mix vehicle routing problem’, Comput. Oper. Res., vol. 18, no. 3, pp. 263-274.
  • Euchi, J., Chabchoub, H. (2010) ‘A hybrid tabu search to solve the heterogeneous fixed fleet vehicle routing problem’, Logist. Res., vol. 2, no. 1, pp. 3-11.
  • Gass, S. (1987) ‘Managing the Modelling Process: A Personal Reflection’, European Journal of Operational Research, vol. 31, no.1, pp. 1-8.
  • Gifford, T., Opicka, T., Sinha, A., Vanden Brink, D., Gifford, A., & Randall, R. (2018) ‘Dispatch Optimization in Bulk Tanker Transport Operations’, Interfaces, vol. 48, no. 5, pp. 403-421.
  • Li, F., Golden, B., Wasil, E. (2007) ‘A record-to-record travel algorithm for solving the heterogeneous fleet vehicle routing problem’, Comput. Oper. Res., vol. 34, no. 9, pp. 2734-2742.
  • Pidd, M. (1999) ‘Just Modelling Through: A Rough Guide to Modelling’, Interfaces, vol. 29, no. 2, pp. 118-132.
  • Pidd, M. (2003) ‘Tools for Thinking: Modelling in Management Science’, 2nd edn., John Wiley & Sons, Chichester, UK.
  • Ritchie, C., Taket, A., Bryant, J. (eds) (2001) ‘Rational Analysis for a Problematic World Revisited’, John Wiley & Sons, Chichester, UK.
  • Schwarz, M. (2003) ‘A Multilevel Analysis of the Strategic Decision Process and the Evaluation of Shared Beliefs’, In: Chakravarthy, B., Gruneter, M.-S., Lorange, P., Lechner, C. (eds) ‘Strategy Process: Shaping the Contours of the Field’, pp. 110-136, Oxford: Blackwell Publishing.
  • Tilanus, C. B. (1985) ‘Failures and Successes of Quantitative Methods in Management’, European Journal of Operational Research, vol. 19, no. 2, pp. 170-175.
  • Williams, T. (2008) ‘Management Science in Practice’, John Wiley & Sons, Chichester, UK.
  • Yaman, H. (2006) ‘Formulations and valid inequalities for the heterogeneous vehicle routing problem’, Math. Programming, vol. 106, no. 2, pp. 365-390.

Leave a Comment