Data and Decision Sciences in Action 2

This paper explores the impact of Vendor managed inventory (VMI) on a decentralized supply chain by proposing two Mixed-integer linear programming (MILP) models for a dynamic lot sizing problem. The models describe the decision-making for lot sizing before and after the implementation of VMI. The proposed models highlight that VMI takes advantage of centralized decision-making, and can reduce the cost of the lot sizing by better synchronization of the decisions. Numerical results, provided to compare the efficiency of VMI with the traditional decentralized lot sizing, indicate significant cost reduction under VMI. A set of experiments is also designed to determine the impact on retailers’ cost (ordering and holding) as well as vendors cost (setup and holding) on the gap.

It is estimated that, at present, around one third of all food produced is lost, either in production and distribution or after retail. To further complicate matters, uncertainty and variability in the commercial and natural environments must also be taken into account when trying to reduce food losses. The objective of the present paper is to develop a decision support system to increase the efficiency of the Australian broccoli supply chain and reduce food losses considering uncertainty. To that end, we develop a two-stage stochastic mixed-integer linear programming model to assist Australian broccoli producers in taking the most cost-effective investment decisions and, at the same time, reduce the losses by producing novel, high value-added products from produce discarded on the field or during transportation. The stochastic model we propose selects, in the first stage, the optimal location of processing facilities to add value to the produce that would otherwise be considered as food loss, and suggests transportation operations as the recourse decisions. The model is solved using Lagrangian decomposition and the subgradient method. The data used to feed the model was collected on the field through a survey applied to broccoli growers nationwide. The model suggests near-optimal investment decisions that are far from the worst possible outcome, had the final market and environmental conditions turned out to be very adverse. Our results represent viable operations for the industry in the medium term.

In this study, we consider a multi-item capacitated lot sizing and scheduling problem for perishable food products, which have a fixed shelf life period due to depreciation issues, such as physical deterioration or perceived value loss. We incorporate shelf life constraints within a classical lot sizing and scheduling problem that considers lot sizing and partial sequencing of production on a single machine over a finite planning horizon associated with demand in each period. The model includes setup times and setup costs. Moreover, it considers that disposal occurs when products reach their shelf life in inventory. We present two variants of this lot sizing and scheduling problem integrating shelf life constraints and also with and without disposal. We test the performance of the proposed formulations on a set of instances from the literature.

The Maximum covering and patrol routing problem (MCPRP) is concerned with the allocation of police patrol cars to accident hotspots on a highway network. A hotspot is represented as a time window at a precise location on the network at which motor vehicle accidents have a high probability of occurring. The nature of these accidents may be due to speeding, driver fatigue or blind-spots at intersections. The presence of police units at hotspots serves as an accident prevention strategy. In many practical applications, the number of available cars cannot cover all of the hotspots on the network. Hence, given a fleet of available units, an optimization problem can be designed which seeks to maximize the amount hotspot coverage. The cars must be routed in such a way as to avoid multiple contributions of the patrol effort to the same hotspot. Each police car is active over a predefined shift, beginning and ending the shift at a fleet station. In this paper, we introduce a method for constructing a time-space network of the MCPRP which is suitable for the application of a branch-and-price solution approach. We propose some large-scale test problems and compare our approach to a state-of-the-art Minimum cost network flow problem (MCNFP) model. We show that our branch-and-price approach can outperform the MCNFP model on selected large-scale networks for small to medium fleet sizes. We also identify problems which are too large for the MCNFP model to solve, but which can be easily handled by our approach.

The vehicle routing problem consists of finding cost-effective routes for fleets of trucks to serve customers. Logistics managers often prefer routes to also be “visually appealing” because of the better flexibility they provide in coping with small alterations, required due to last-minute or unforeseen events. Compactness of the routes is a key desirable feature, and it can be accomplished by minimizing the area enclosed by the routes. A common approach in the literature relies on imposing a penalty on the area of the convex hull. We propose to use new features which are well correlated with the convex hull area but are significantly easier to implement, having O(n) computational complexity instead of \(O(n \hbox {log} n)\). By accepting only a minimal loss of quality with respect to a primary objective function, like the routes’ total length, we show that area-type penalties can be effective in providing good guidance: construction methods which are based on insertion are naturally steered towards routes displaying more attractive shapes. Used in conjunction with an adaptive large neighbourhood search, our new proposed features lead to routes that exhibit similar compactness compared to using a convex hull area penalty. We also achieve good separation between routes.

When managing inventory or supply systems, it is important to make good choices about which stock to prioritize over others. We can improve the overall availability of the supplied systems by making optimal choices on which inventory items should be allocated to meet demands. In this paper, we will show how machine learning algorithms can be used to prioritize inventory. The developed algorithms were tested on a real data set and the improvement in inventory allocation measured. Machine learning is a powerful technique for transforming inputs to outputs in order to best achieve a set goal. It has many applications in areas where there is an abundance of data, and where the resulting decisions can be measured. As such inventory management is a suitable area of application, in particular the prioritization of supply. Such an approach is even more relevant to those inventory models that represent autonomous processes. The models we are interested in are those relating to system availability and that use item backorder calculations. These models rely on a traditional prioritization approach known as marginal analysis, otherwise known as a process of marginal allocation using a greedy algorithm. Because marginal analysis does not take into account performance over time, nor complex relationships in data sets, there may be potential for a machine learning algorithm to provide better results if it can learn to exploit both temporal and relationship data. The benefit of such an improvement is the value of availability generated and cost savings made in the supply network.

We study a capacity alignment planning problem for a coal chain. Given a set of train operators, a set of train paths and a terminal comprising of a dump station and a set of routes from the dump station to the stockyard, we seek a feasible assignment of train operators to train paths, to time slots at the dump station, and to routes. The assignment must maximize the number of system paths in the resulting schedule and the schedule should perform well with respect to various performance criteria. We model the problem as a mixed-integer conic program (MICP) with multiple objectives which we solve using a hierarchical optimization procedure. In each stage of this procedure, we solve a single objective MICP. Depending upon whether we evaluate the associated performance criteria under a 2- or 1-norm, we reformulate the MICP as either a mixed-integer second-order cone program or as a mixed-integer linear program, respectively, and can streamline the hierarchical optimization procedure by exploiting properties of the model or observed behaviour on practical instances. We compare the performance of the procedure under the different norms on a real instance of the problem and find that the quality of the solutions found by the faster 1-norm procedure compares well to the solution found under the 2-norm.

Aerial firefighting resources are an integral part of modern wildfire suppression strategies. In many locations around the world where wildfires pose a serious threat, firefighting authorities have access to fleets of different aircraft. These can be used to provide support to land-based resources during the extended attack of existing fires or to quickly suppress recent spark events during the initial attack phase. As the amount of time that a fire has been burning is a predictor of the amount of damage it causes, fast aerial response times are critical. Therefore, there is significant value in dynamically repositioning aircraft to airbases and fires over the course of a fire day or fire season. In this paper, we devise one such approach based on model-predictive control to make relocation decisions at various times over a single day. These relocation decisions are based on solving an underlying Mixed-integer linear program (MILP) so as to minimize expected damage over a lookahead horizon. The inputs to this program are updated at each of these decision times based on prevailing stochastic weather conditions, the current state of fires in the region, and the current assignment of aircraft to bases and fires. The expected fire damage profiles used in this model are based on empirical data that is pre-computed for the region of interest. We apply our model to a scenario in Central Chile and show that with careful parameter selections it is possible to make improved relocation decisions to reduce the expected fire damage in a region using this approach.

An efficient operating theatre schedule contributes significantly to enhancing the efficiency of hospital operation management and plays a critical financial role in most hospital settings. In this paper, an operating room scheduling problem based on patient priority is investigated at tactical and operational levels subject to specific strategic decisions. At the tactical level, the main goal is to generate a cyclic time table, known as the master surgical schedule (MSS) and can be repeated over the planning horizon of several months to years. Operational level concerns about allocating patients to operating rooms and determining the day of surgeries, which is called the surgical case assignment problem (SCAP). To handle the problems at both decision levels simultaneously, known as the MSS-SCAP problem, an integer linear programming (ILP) model, called MSS-SCAP model, and a heuristic approach are proposed. The objective function is to maximize the total priority scores of the patients assigned to the surgical scheduling blocks over a given planning horizon. An adaptive ILP model is also proposed to solve the SCAP, taking into consideration the dynamics of the waiting list. The computational experiments are conducted using a set of random data to evaluate the performance of the proposed MSS-SCAP model and heuristic algorithm, in terms of solution quality and computation time. Our numerical results indicate that the proposed ILP is capable of yielded optimal solutions for the small-scale instances and near-optimal solutions for medium-size instances within 3,600 seconds. The proposed heuristic algorithm can generate quality solutions within 2 seconds for large-scale instances.

A fantasy sport competition is an online competition in which participants act as the coach and selector of their own fantasy team of real players. These competitions are remarkably popular with currently over 5.5 million teams in the major competition of the English Premier League. A fantasy team scores points based on the statistical performances of the team’s players in their real-world sporting matches. The objective for each coach is to finish the season with the highest total number of points. During the season, coaches must manage a budget as well as trade players in and out of the team subject to a number of constraints. Due to their well-defined nature, as well as the simple objective function, these competitions lend themselves very naturally to analysis by Mixed Integer Programming (MIP). In this paper, we consider three different problems for the 2018 season of the AFL SuperCoach competition, modelling and solving each with MIP. The aim of each problem is to highlight the gap between what was achieved by real players and what was theoretically possible. The first problem is to determine all the decisions that a coach should have made to obtain the highest score possible. The second problem is to determine the lowest starting budget from which it would have been possible to win the competition. The third problem is to determine whether it would have been possible for a team that was set up at the start of the competition and completely forgotten about to win the competition.

Contemporary risk management methodologies are typically used for identification and prioritisation of strategic risks. The International Risk Management standard, ISO 31000:2009, is the world-wide basis for best practice in strategic level risk processes. However, due to the qualitative and subjective nature of strategic risk, its analysis requires a more nuanced approach than that used in more tactical or operational settings and this paper discusses the need to understand the range and nature of strategic threats, and how to represent risk assessments. As such, a particular focus of this work is on how to incorporate best practices in strategic risk analysis, and operations research into the design and application of strategic risk management in the Defence context. A number of steps are recommended incorporating international risk management best practices within the context and uncertainties unique to strategic risk management for Defence (as opposed to tactical or engineering risk management).

Without liquid petroleum, either jet fuel, diesel, or other products such as lubricants and fuel oil, the Defence force ceases to function. Furthermore, the national support base, the bed-rock of the Defence force stops as well. Petroleum products will continue to form an energy source of choice for Defence because of superior energy density for decades to come. While Australia is seen as a regional power, its energy resilience is in a state of change and is generally seen as declining. Currently the Government sees market forces as providing petroleum supply security, given the nation stores 50–55 days of stocks which is below the mandatory 90 days required by the International Energy Agency. With this context, the Defence Science and Technology (DST) Group has developed a security-based global petroleum simulation, called SPECULA, in order to model the effects of regional and global changes in oil production, refining, shipping or distribution of petroleum products during conflict or significant environmental events. SPECULA is a simulation model, where petroleum is transported globally, based on regional variations in price. Price here is modelled as a “pseudo-price” which has a global component and a regional component. The global component is based on the difference between global supply and demand. The regional component is based on the regional inventory level. If the inventory level is low, the regional price rises. Trading, or the movement of petroleum from one region to another occurs because of inter-regional price differences, commonly known as arbitrage. The SPECULA model is spatial, as tankers move cargos along inter-regional seaborne routes. This paper briefly describes the security context of the petroleum supply chain in the Asian region. Then previous economic models of Australia’s petroleum supply security are reviewed and critiqued in terms of their ability to model conflict scenarios. The SPECULA model is then described along with model parameters and outputs. Finally, the future challenges of this model are addressed in the discussion/conclusion.

This paper describes a methodology for reviewing organisational-level adaptability from a systems perspective. Taking an action learning approach, we reviewed the Australian Army Lessons Network (ALN) as a case study in order to: (i) develop practical options for improvement of the ALN, (ii) reflect on the review methodology and identify options for improving its effectiveness in subsequent reviews, and (iii) demonstrate the utility of adaptive review and lay the foundations for its further application.

Land Combat Vehicles (LCVs) are a critical fighting capability of the Australian Army. The operational effectiveness of a LCV is usually modelled via combat simulation in which the multi-criteria metrics are measured from the simulation output. Consequently, it is important to develop a multi-criteria decision-making procedure to support upcoming acquisition decisions for future vehicle options. Criteria measurements in combat simulation and decision-makers’ preference often involve uncertainties; however, option ranking and selection procedures from simulations are normally limited to a single response metric or deterministic preference for the multiple metrics in the current literature. In this paper, we address these uncertainties by using a probability distribution function and Monte Carlo simulation in the stochastic multi-criteria acceptability analysis (SMAA) model for aiding this decision-making problem. Additionally, all uncertain preference information from DMs are represented as feasible weight space (FWS) and are used in combination with other weighting techniques such as analytical hierarchy process (AHP). The aim of this paper is to describe the application of SMAA, FWS and AHP to the results generated in a close-loop combat simulation, such that the options with uncertain data can be evaluated and analyzed, and the best option can be selected for a specific task or scenario. To the best of our knowledge, this combined approach has been applied for the first time to deal with the defence decision analysis problems with uncertainty and interdependency.

In armoured fighting vehicle design, the Iron Triangle concept describes the design tensions that exist between the three primary characteristics of these vehicles: mobility, protection and lethality. Traditionally, wheels and tracks represent two different trade-off instances between different aspects of these three factors and are suited to different operational conditions. To provide some clarity to the wheels vs tracks argument for the ADF, a wheels vs tracks study was undertaken in the Australian context. This study collated results of previous studies and performed a meta-analysis, synthesizing the results to produce an understanding of the impacts of wheels and tracks on operational outcomes, analyzing the current evidence of the strengths and weaknesses of wheels and tracks, and interprets these in different contexts characterized by environmental and operational variables. The results of the meta-analysis show that overall a tracked vehicle will offer a greater operational capability advantage more often. Out of the 72 different contexts defined, 62 show an operational advantage for tracked vehicles. Only nine contexts had an overall utility skewed towards a wheeled vehicle, and in one context wheeled and tracked vehicles were judged as equal. The analysis identified 14 contexts with an intensity rating of extreme, and in all of those contexts tracked vehicles had an operational advantage over wheeled vehicles. In 15 of the 20 contexts judged to be most likely, tracked vehicles had an operational advantage over wheeled vehicles, while the remaining five showed an operational advantage for wheeled vehicles.

A modern Defence Force consists of a diverse range of capabilities to support missions at the tactical, operational and strategic levels. Designing a balanced and affordable force structure to meet Government strategic objectives and assure national security has always been a challenge. Force design is a centralized and enduring process within the Australian Defence Organisation that seeks to translate Government strategic objectives into a coherent force structure within specified time and budget envelopes. This process increasingly relies on analytical approaches and tools such as wargaming, simulation and optimization techniques. This paper investigates evolutionary algorithms (EAs) as a potential tool for generating and evaluating force structure options. EAs can evaluate an extremely large solution space of force mixes at a much faster rate than human cognition to determine a balanced and affordable force structure option according to an objective function. This paper also discusses the implementation of a software framework, dubbed “FORCESIGHT”, which can be customized by developers to model any scenario where the use of EAs is appropriate. Based on the outcomes of a trial of FORCESIGHT, it is clear that the EAs approach could provide a result of respectable quality. It is demonstrated that EAs can lead to large increases in efficient evaluation of potential improvements to the Force-in-Being and Future Force.

Behaviour trees offer a means to systematically decompose a behaviour into a set of steps within a tree structure. Genetic programming, which has at its core the evolution of tree-like structures, thus presents an ideal tool to identify novel behaviour patterns that emerge when the algorithm is guided by a set fitness function. In this paper, we present our framework for novel behaviour discovery using evolved behaviour trees, with some examples from the beyond-visual range air combat domain where distinct strategies emerge in response to modelling the effects of electronic warfare.

Random Forests is considered one of the best off-the-shelf algorithms for data mining. However, it suffers from poor interpretability and an opaque decision structure. In this paper, we develop a method for generating an “expanded basis set” for a Random Forest model that captures every possible decision rule and vastly improves the transparency of the classifier. The expanded basis set allows the structure of a Random Forest model to be algebraically manipulated and facilitates a number of operations, including inverse mapping from outputs to the domain of inputs, systematic identification of every decision boundary, and comparison of Random Forest models. The expanded basis set facilitates visualization of the global behaviour of a Random Forest classifier and a data set by combining parallel coordinates with a non-linear binning transformation. The global visualization allows classifier performance to be compared against domain expertise, and areas of underfitting and overfitting to be readily identified. Additionally, the expanded basis set underpins the generation of counterfactuals and anchors—combinations of variables that control the local outputs of a Random Forest model. The basis states can also be used to place bounds on the model stability in response to single or multi-feature perturbations. These stability bounds are especially useful when the model inputs may be uncertain or subject to variation over time.

Analysis of output data from closed-loop combat simulations can provide insights into the relationships between model inputs and outputs. However, analyses that only consider end-of-run output data may not be sufficient to explain why those relationships exist. In the case of stochastic models, where multiple replications of the same scenario are conducted, the presence of outliers and multi-modal results also needs to be accounted for. In this paper, we use a military case study to explore a range of techniques to interrogate the intra-run event data (i.e. trace logs) generated by a combat simulation in order to help address these two issues. Cumulative event plots and geo-spatial visualization techniques which also incorporate the temporal aspects of the simulation appear best suited to explain the presence of outlier replications and multi-modal results. Exploratory work using hierarchical clustering techniques and temporal decision trees provide a promising step towards better explaining causal events within the combat simulation data.