Improving customer demand and cost forecasting methods |
Technical Impact
Improving customer demand and cost forecasting methods
This collaborative R&D project aims to improve the abilities of inventory and capacity forecasting models through the development of Artificial Intelligence (AI). These tools will be able to autonomously identify, search for, collect and analyse large amounts of data from a wide variety of disparate data sources. The development of these tools will require the selection, development and integration of individual methods and techniques that include neural networks, fuzzy logic, genetic algorithms, ontologies and semantic networks. The overall system to be developed is illustrated in Figure 1 and the technical approaches to be used are as follows:
Development of generic Inventory and Capacity Forecasting Knowledge Identification Systems (KISs) - WP1 & 2:
The complexity and diversity of the inventory and capacity forecasting domain knowledge and terminology is a major hurdle for the successful development of the AI methodologies. Fortunately, such hurdles have been overcome by the emerging ontology technology and the establishment of ontologies for both inventory and capacity forecasting will be the first step towards autonomously identifying the knowledge in this domain. The initial step in building the ontology will be to identify the purposes, scope, and requirements of the ontologies. Based on these requirements data and information about inventory and capacity forecasting will be collected using a variety of sources including published research, other ontologies and importantly experts from partners ULL, Trelleborg and Preactor. The ontology must maintain predictor-predictor variable and predictor-dependent variable relationships. Using these relationships the ontology will support knowledge collection by identifying potential data types and data sources and providing navigation paths during data collection. In doing so the KIS must analyse users requests in terms of model characteristics such as levels of the forecasting accuracy required.
Development of Industry Specific Knowledge Collection Systems (KCSs) for inventory and capacity forecasting in the automotive industry, capacity forecasting of sustainable energy and inventory forecasting in the general engineering sector - WPs 3a-3d:
The purpose of these KCSs will be to undertake the actual search, navigation and aggregation processes required to find appropriate data sources and extract from these sources the individual items of data that may represent predictor variables and/or dependent variables within the resulting forecasting models. Fuzzy logic will be used to undertake a ‘rough cut’ analysis of each data source as a means of reducing the amount of data moving on to the data analysis stage, ie it will make subjective decisions as to whether individual data types have causal relationships with capacity/inventory demand.
Development of Generic Inventory and Capacity Forecasting Knowledge Analysis Systems (KAS) - WPs 4a & 4b:
The data and knowledge collected from the KCSs will be maintained in an appropriate database. Methods will then be developed for analyzing this data in order to discover appropriate relationships from which forecasting models can be developed. Models will be required at different decision making levels with organizations, ie strategic, tactical and operational, as well as for use in differing time periods, eg annually, monthly, daily. Professor Stockton’s prior research which addressed the cost model development process will be extended to establish appropriate model characteristics, such as level of accuracy and detail, which will guide the data analysis process. The application of neural networks and/or fuzzy logic will enable a wide range of models to be developed. For example, where strategic forecasting is required of new products and little data was available then fuzzy logic would be used. In complex situations, where there were large numbers of variables and large quantities of data then the system would automatically select an artificial neural networks approach. WP4c will attempt to combine the two application specific KASs into a single KAS.
Development of knowledge systems optimisation routines - WP 5:
In order to ensure that the system is as far as possible autonomous the KISs, KCSs and KASs will form part of a closed-loop control feedback system. The forecasting error generated by the models output from the KASs will be used to provide the fitness function values for a GA Optimiser. The optimizer will have full control over the structure and content of the KICs, KCSs and KASs such that they can be modified during the GA evolutionary optimisation processes.
WP6a to 12:
These remaining work packages, ie, are essential to successful project management, product control, and validation and exploitation of the forecasting tool set. However, they contain little novel or innovative technical or scientific input and have not, therefore, been included within this Technical Approach description.
Technological Functions of the AI Components
An overview of the relevance of the main AI techniques that will be used is provided.
- Fuzzy Logic (FL), will be used to model the world in a way that is representative of the uncertainty inherent in data, language and thought, i.e. it will allow us to both represent and model vagueness imprecision and uncertainty. In this way we can model human expertise using linguistic variables. The approach taken, therefore, is to represent the uncertainty in the values of each variable type using fuzzy sets, combine these fuzzy sets using rules, inference with the rules and then use defuzzification to produce a final value. Professor Bob John, DMU, will provide the scientific and application expertise in this area.
- Artificial Neural Networks (ANNs), will allow us to develop networks that take known inputs, (i.e. variable values), to produce outputs, (e.g. values of sales demand). Depending on the application the ANNs will be ‘trained’ using known or unknown outputs to learn relationships between predictor variables and the sales data. ANNs are particularly good at pattern recognition and hence are able to discover relationships between data types. Dr Martin Ziarati, CfFF, will provide the scientific and application expertise in this area.
- Genetic Algorithms (GAs), are an ‘evolutionary computing’ based optimisation technique that borrows from the biological notions of genetics and natural selection. Essentially their probabilistic approach to searching large spaces will be used to optimise the Fuzzy Logic and ANN modelling procedures as well as the model forecasting relationships with the objective of optimising forecasting accuracy. Professor David Stockton, DMU, will provide the scientific and application expertise in this area.
- Ontological Approaches combined with the use of Semantic Networks will be adopted for the development of processes for searching, aggregating and deciding which data should be collected. The approach to the project is one of sharing knowledge between different sectors, with automotive, sustainable energy and general engineering being represented in the project partnership. The knowledge contained in these important sectors when forecasting operating resources is to be combined to produce an ontology most suited to this shared knowledge. The ontology will allow for a specification of generic concepts and relationships that will assist inventory and capacity forecasting. Professor Bob John, DMU, will provide the scientific and application expertise in this area.
Roles of Industrial Partners:
In terms of the Technological Approach the industrial partners will assist in identifying the basic building blocks from which application specific forecasting application ontologies are constructed, mapping of the supply chain forecasting environments within the partner organisations, and identifying the potential forecasting model parameter variables and data sources for these variables.