Production scheduling has many features in common with transport scheduling. Infomace International has a proven record of milk tanker scheduling in both New Zealand and Australia. Skills and experience gained in transport scheduling are immediately applicable to production scheduling.
The basic elements of
are common to both disciplines. Some of the simplest production scheduling problems turn out to be mathematically equivalent to the famous Travelling Salesman Problem, the most basic of transport scheduling problems.
It is well known that production scheduling problems are mathematically very difficult. A great deal of research has been done in this area, and it has been proven that optimal scheduling for even a single machine is a very complex matter. In can be done, but you never know until too late just how long it is going to take.
Fortunately, there is a wealth of techniques available to produce good schedules, even if perfect schedules remain elusive.
A heuristic is a rule of thumb, a simple rule that allows decisions to be made quickly and without much thought. A well-chosen heuristic can produce very good results for very little outlay.
Mize, White and Brooks list seven common production scheduling heuristics:
Conway found 92 different rules that had been used or seriously considered - and that was in 1964!
If a well-chosen heuristic is a boon, a badly chosen heuristic can be a disaster. For example, under shortest operation first, long operations may never get done at all. If a company relies on heuristics to do its scheduling, it should conduct regular audits to see that the rules are still appropriate.
Heuristics can be analysed theoretically (eg. Chen 1995) or (more usefully) by simulation. A simulator is a computer model that applies each rule to a set of historical data, and compares the outcomes on the basis of standard measures (lateness, change-over time etc.). This allows direct and objective comparison of heuristics, so that the best heuristic can be chosen. Infomace International has used simulation to audit its own vehicle scheduling system.
A classic study of dispatch rules by Earl LeGrande (1963) evaluated six rules at Hughes Aircraft in El Segundo California. The Hughes fabrication shop had 1000 machines, 400 to 500 workers and 1800 to 2500 orders in process; it completed 100 to 150 shop orders per day. The performance of the rules was evaluated by simulating the shop in detail. The "minimum processing time" rule had the best efficiency and flow rate, while the "minimum slack time per operation" rule was best at meeting due dates. The rules "first come first served" and "earliest planned start date" did surprisingly poorly - worse than "random", in fact.
When dealing with complex problems like transport or production scheduling, Infomace has found that the best approach is to use improvement methods. Given a measurable objective (essential for all optimisation work!), it is possible to tell which of two rival schedules is better. If Infomace also has a set of techniques for modifying existing schedules, then Infomace can try them in turn and choose the best modified schedule. Infomace now uses the modified schedule as the basis for further improvements. And so on.
This is a good strategy because the schedule is always getting better, and Infomace always has a workable schedule. This means Infomace can stop whenever the schedule is "good enough", or no further progress is being made, or if Infomace has run out of time and has to implement whatever Infomace has got.
Another advantage of this approach is that many different styles of improvement can be used together, leading to better schedules than any one method can produce alone. Multiple improvement methods produce synergy, but multiple heuristics do not.
Of course, Infomace has to start with a workable schedule, but there are plenty of heuristics available. A good improvement suite can do a great deal with a mediocre initial heuristic.
Modern manufacturing theory (WCM, JIT) aims to solve scheduling and inventory problems by completely restructuring the plant. Machine positioning, procedures, personnel, management structures and planning are completely overhauled. Scheduling is made easier by short lead times and internal inventory is abolished completely. The result is effectively a completely new factory and a completely new company. Such fundamental restructuring requires enormous courage and commitment at all levels within the company, and full cooperation from the trade unions (Hay 1988, Schonberger 1986).
In statistical parlance, demand for a product constitutes a time series: a quantity that depends on time in a way that is only partly predictable.
There are two traditional approaches to predicting time series. Firstly, historical data can be decomposed into trend (long term behaviour), seasonal cycles and residuals (random effects that cannot be predicted from historical data alone).
Secondly, future demand can be estimated as a weighted moving average of past demand. Infomace International has used this second approach to predict milk supply, producing more accurate predictions than the dairy company had previously obtained. More complex models integrate both these techniques.
Finally, the product demand need not be treated in isolation. If it is thought that interest rates or other economic indicators may have an effect on demand, this can also be analysed and used to improve the accuracy of forecasts. This can be approached in a structured manner (causal forecasting) or by searching for empirical relationships in apparently unconnected data (data mining). In either case, the statistical method is known as regression analysis.
Armstrong and Grohman (1972) used causal forecasting to predict total demand for US air travel as the product of seven factors raised to various powers. The most important factors were current demand, US population, ticket price, and a general increase of 12% a year. Less important were aircraft speed, the safety record and the state of the economy. Armstrong and Grohman found that their causal model was a more accurate predictor than time series analysis of sales figures alone. As an added bonus, their model supplied strategic information and not just sales predictions.
An example of data mining is the oft-quoted connection between share prices and sunspot activity. There is a good statistical fit, but no known mechanism whereby sun spots affect the share market (or vice versa!). For this reason, data mining must be treated with caution.
When the number of variables becomes very large, statistical methods become unwieldy. Neural networks offer an alternative approach when many variables interact in complex ways. Neural networks are modelled on the human brain, and are capable of being trained and learning on the job. Infomace International has experience with neural networks.
Neural networks can also be used for process control. Many production systems have complex behaviour which is not easily unravelled by detailed analysis. Neural networks can take a holistic view and can reveal (and respond to) large scale features that might otherwise be missed.
Other proven applications include real-time adaptive image compression, industrial inspection, MICR digit recognition, estimate probability functions, statistical, financial forecasting, credit rate scoring, scheduling (routing), moulding plastics, fraud detection, weather prediction, process control, targeted marketing, multi-spectral image processing. There will be many others not covered here. One is limited only by one's imagination.
Business logistics is a discipline which addresses issues concerning the efficient and cost-effective flow and storage of products from the manufacturing sources to the customers. The goal in a logistics system is to maximise profits and provide a least total cost system, while achieving desired customer service levels.
Computer models have been used in logistics planning for over thirty years. For example, models have been developed to determine the best locations of production and distribution facilities to minimise transportation costs subject to meeting pre-specified delivery time constraints.
The following list gives a sample of computer applications in business logistics:
Most advances in the use of computer-based systems to support logistics have been in the areas of transportation and low level operational decisions, while little attention has been applied to the areas of product forecasting and strategic support systems. The effective management of distribution activities may be achieved by those strategic decisions which attempt to minimise inventory and transportation related costs. This means that there is a need to identify the optimum quantity and location of distribution nodes (e.g. warehouses and distribution centers), and the most cost-efficient distribution channel for each type of customer order.
Many existing models focus on individual components of the overall system, and thus ignore the integrated approach. An integrated approach to distribution planning is essential due to the inherent trade-offs involved. Examples of such trade-offs include customer service levels versus inventory carrying costs versus transportation times. Traditional models typically fail to approach the problem from an integrated marketing concept. They include only logistics-related costs but ignore selling costs, promotions, order processing costs, etc.
Most models include only costs and therefore cannot quantify the impact on corporate profitability of various alternative scenarios. Also, a major limitation of these systems is that they are not user-friendly.
Most of the tasks required for effective management of logistics and distribution activities can be achieved using a computer-based decision support system described in the next section.
A Decision Support System (DSS) is a computer-based information system designed to assist both managers and analysts in the process of decision making through the use of sophisticated software technologies, operations research and management science models. Such a system must provide improved effectiveness in the decision making process, and the capability to analyse complex distribution networks and marketing environments within a short time frame.
Decision support systems provide immediate access and flexible analysis of data, including access to a variety of databases and models as required. The components of such systems include database management systems, and facilities for interacting with the users. Decision makers need to estimate consequences of proposed actions and model situations for finding optimal solutions.
A decision support system used in logistics should be capable of performing the following:
Operations research / management science models provide the theoretical foundation for the construction of a decision support system.
Major approaches regarding distribution modeling are:
Optimisation via such mathematical programming techniques as linear, non-linear, and integer programming should be employed whenever possible since they guarantee the best feasible solution.
The complexity of modern logistic systems, and their resulting data and calculation requirements, however, usually demand extensive computer hardware capabilities and prohibitively long processing times. Heuristics reduce the search space to a manageable number of feasible alternatives which can then be analysed by one of the other approaches. Heuristics have been applied to a variety of logistics problems including routing, facility location, scheduling, and logistics system design.
Simulation is a mathematical description of a decision problem, usually in sufficient detail. This type of mathematical modeling is intended to replicate the dynamics of an existing or planned system rather than attempting to find a feasible solution. Due to the level of detail required in logistics system design, the simulation approach is frequently recommended.
While the objective of reducing costs and satisfying the demand for service remains the same, the technologies to realise these objectives change rapidly. The relevant technologies are those which mimic different aspects of intelligence and can be used in the complex task of decision making. The next section investigates a number of such technologies under the title of computational intelligence.
Intelligent systems emulate human mental faculties such as adaptation and learning, planning under large uncertainty, coping with large amounts of data, etc. Successful industrial applications of intelligent systems usually deal with several of these aspects. Therefore, it is natural to combine various technologies with different capabilities within one system.
Hybrid intelligent systems address these problems using the fields of Artificial Intelligence (AI) and Computational Intelligence (CI). The field of AI and especially its successful application “expert systems” are characterised by symbolic processing, rule bases and representation of knowledge. CI systems, on the other hand, are collection of methodologies like neural networks, fuzzy logic, and genetic algorithms which are concerned with numerical processing of large amounts of data, learning, and adaptation. Simulation studies can be effective in the integration of computational intelligence into systems. Reducing subjective decisions and increasing the potential for real-time automation are practical goals for such projects.
Expert systems perform reasoning using pre-established rules for a well-defined and narrow domain. They combine knowledge bases of rules and domain-specific facts with information from users about instances of problems. For applications with well-defined rules, expert systems can be easily developed to provide good performance. Furthermore, most development software packages allow the creation of explanations to help the user understand questions being asked. A major limitation of the expert system approach arises from the fact that experts do not always think in terms of rules.
An artificial neural network is a model that emulates a biological neural network. The nodes in a neural network are based on the simplistic mathematical representation of what Infomace thinks real neurons look like. Most of today’s neural networks are simulations of massively parallel processes involving processing elements in a highly interconnected architecture.
Neural networks rely on training data to program the systems. Particular applications are developed by establishing an appropriate training set that allows the system to learn and generalise for operation on future data. Inputs that exactly match the training data are recognised and identified, while new data (or incomplete and noisy versions of training data) can be put into closest matches to patterns recognised by the system.
Neural networks can be preferable to expert systems when rules are not known, either because the topic is too complex or no human expert is available. If training data can be generated, the system may be able to learn enough information to function as well as, or better than, an expert system. The data-driven property of neural networks allows adjustment of changing environments and events. Another advantage of neural networks is the speed of operation after the network is trained; the natural use of parallel systems and neural chips enhances this aspect dramatically.
Fuzzy systems mimic an aspect of human reasoning by doing approximate reasoning. They can provide an interface that seems more like the interactions human decision makers experience with each other. This technique can broaden the usefulness of expert systems and neural networks, allowing operation in grey areas where precise values may not be known or may not be necessary to draw conclusions.
Modeling after concepts of biological evolution, the genetic algorithm technique provides a search technique that is an alternative to those in traditional AI. It uses the aspect of learning from experience and the use of information to identify inferior or preferred solutions. Thus, GA components can enrich a hybrid computational intelligence system and raise the level of intelligence.
Application Of Computational Intelligence In Decision Support Systems
An important area of information systems deals with tools and techniques that aid decision makers, especially at the middle and top levels of management. The domain of applicability includes dynamic, open systems subject to considerable uncertainty and risk where problems tend not to be well structured, and exact solutions and data requirements are difficult to anticipate.
Neural networks can provide capabilities not available in decision support systems or expert systems, specifically, the ability to adapt to new situations typical of open systems and to generalise from experience and to interpolate from learned facts to recognise similar situations.
A natural application of neural networks is the processing of large data sets to identify patterns and features that require further attention and may reveal the need for decisions. Neural networks could be components of database mining systems that run in the background to look for problems or interesting correlations in a database that may be of interest to a managerial decision maker.
A goal for intelligent database systems is to handle information and decision making in a way more similar to humans, and the neural network components may be crucial for finding patterns in data, finding approximate matches and best guess estimates, and facilitating inexact queries.
The new generation of decision support systems which are currently developed for business logistics take advantage of the powerful features of artificial and computational intelligence methodologies. Thus, the new systems will have more advanced features than the traditional ones.
Some of these features are listed below: