The world around us is a complex and dynamic system. As a result, we need intelligent systems that can learn and adapt in a changing environment without human intervention in order to solve complex and dynamic real-world problems. Cercia is an international leader on these topics and has in-house expertise in solving real world problems using these techniques for many years.
Cercia investigated algorithms that can prioritise task lists and allocate individual tasks to vehicles and personnel. For efficiency, the solutions should distribute new work items on an on-demand basis, taking into account all the relevant constraints. This process needs to be extremely rapid, with the interval between independent task completions potentially measured in seconds. Individual staff should receive a new task instantaneously, when requested. The work distribution should be adaptive, able to cope with warehouse configuration changes such as incorrectly-recorded (or damaged) stock; new personnel arriving for work, or existing personnel leaving; vehicles breaking down; job prioritisation changes; new incoming stock; new purchase orders; updates to the warehouse state.
Financial planning and forecasting are essential for any business. Turning "gut feeling" into sound decision-making based on real data, is a huge challenge for both large corporations and SMEs, as well informed decisions mean profits. Severn Trent Water Ltd. has recently turned to the Centre of Excellence for Research in Computational Intelligence and Applications (Cercia) at the University of Birmingham for state-of-the-art solutions to emerging business challenges. Two phases of a software project in financial planning and forecasting were completed in early 2003 and 2005, respectively. This software has been in use by Severn Trent Water since early 2003 and a major upgrade was completed in early 2005.
Early diagnosis of many diseases is crucial to the successful treatment of an illness. However, it is often very difficult to diagnose a disease correctly based on symptoms and test results. We have developed a very successful evolutionary artificial neural network system, i.e., EPNet, for assisting a doctor to diagnose diseases, including heart disease, diabetes, breast cancer, and thyroid [1,2,3]. The system was tested on real-world data sets and compared against other similar systems [1,2,3]. The results showed that our system performed significantly better than others [1,2,3].
Although we have only applied our system to four types of diseases, it can be applied to other diseases. The key idea behind our system is to use a special evolutionary programming algorithm to evolve both the architecture and weights of the artificial neural network automatically [1,2,3].
The same principle and techniques can also be adapted to the diagnosis of other faulty systems, e.g., engines, manufacturing systems, etc.
Environmental monitoring and protection are crucial to the quality of our life and the ecological system in the world. For example, a major outbreak of the blue-green algae in a fresh water lake can have a devastating impact on the fish and other lifes in the lake.
We have used evolutionary artificial neural networks to predict the activities of Chlorophyll-a in a lake in Japan [4]. We were able to produce much more accurate prediction than other methods [4]. The more accurate prediction would enable us to take an appropriate preventive action and reduce the risk of a major outbreak.
In a telecommunications network, it is very important to know when the network is busy and when it is not. Such information will enable a carrier (company) to make an informed decision on the necessary capacity between two cities as well as set a pricing policy that encourages off-peak use of certain lines. However, it is very difficult to predict traffic flow in a telecommunications network. We have used our newly developed neural network ensembles to predict traffic flow in an Austrian telecommunications network among 32 regions [5]. We have shown that our negative correlation learning algorithm can train an ensemble successfully to solve this problem.
Although the project was carried out for a telecommunications network, the techniques we used can equally be applied to flow prediction in an electricity network, a water network, or a gas network. We will be able to predict the demand within a period based on historical data.
Credit cards are widely used in the world, even in developing countries like China and India. However, issuing a credit card carries some risk because the card holder may or may not pay back the money. It is important for any bank to assess the risk of a potential card holder before issuing a card. We have applied our evolutionary artificial neural networks and neural network ensembles for credit card application assessment in an Australian bank [2,6,7]. Excellent results have been obtained in comparison with other existing methods.
The techniques we have developed for the Australian credit card problem can be applied to other problems, such as insurance fraud detection [8], risk assessment for loans, premium setting for insurance, etc.
The ability to analyse and predict the movement of financial time series is crucial to the success of any financial institution. We have applied our evolutionary artificial neural networks to the analysis of Hang Seng Stock Index and obtained impressive results [9]. More work can be done on portfolio optimisation, bond price prediction, share price analysis, etc.
Many real world problem involves object/pattern recognition, e.g., automatic inspection of products through photos, videos or infrared, target recognition in an image, finger print recognition, number plate recognition, etc. This kind of problems usually requires a recognition system that can deal with a large number of inputs. However, not all inputs may be important to the recognition task. It is essential to extract only relevant features out of a vast number of possible inputs so that a high accuracy recognition system can be constructed. We have invented (EU patent pending) a novel method for automatic feature extraction and object recognition based on neural network ensembles [10], which can be applied to a wide range of real-world recognition problems [11].
Many manufacturing and processing systems involve sensors, motors and robots that must be controlled to operate in a smooth and efficient manner. Such controllers are frequently required to adapt to changing environmental conditions, plant wear, and such like. Rather than attempting to design these systems by hand, one can set up sufficiently general neural network or traditional control systems, and let them learn from appropriate training data sets how best to operate. Such systems can be trained to perform well according to a range of different performance criteria [19]. Even more powerful systems can be developed by allowing whole populations of these control systems to evolve by a process of simulated natural selection [20, 21].
A common real world problem is the need to recognise patterns in data, and use them as a basis for classification. Standard neural network systems are well known to be able to do this well. More specialist neural network systems can be tailored to cope with ambiguous training data [22], and various types of noisy training data [23]. Further insights into the data and the tasks under consideration can be obtained by analysing the internal representations learnt by the neural networks [24]. Complex multi-task problems, such as the requirement to perform two or more distinct classifications at once, may benefit from building some form of modularity into the system, and evolutionary strategies can be employed to optimise these more complex architectures [25]. Similar evolutionary approaches can also be used to optimise the speed of learning in neural network systems [26], which will be particularly important when real-time training is required.
Design and optimisation are ubiquitous, from circuit design to toy design, from turbine blade design to truss design, from inventory minimisation for a company to wastage minimisation in stock cutting, etc. Computational intelligence techniques have proven to be extremely competitive in solving complex real-world problems in comparison with existing mathematical programming methods. Cercia is at the forefont of applying computational intelligence techniques in real world design and optimisation problems.
"Nature Inspired Creative Design" is a new research network, part of the AHRB/EPSRC funded "Designing for the 21st Century" program. Cercia is creating a cluster of scientists, artists, designers, and industrialists that focusses on taking ideas, methods, paradigms and algorithms from nature and introducing them into the design process. Nature inspired approaches have the potential to create better designs, as well as better design processes and tools.
Stock cutting is a problem faced by many workers in the world. For example, when steel pipes are manufactured, they are produced in certain fixed length. However, customers normally have different requests for pipes of different length. To save costs, we need to minimise wasted remainders. We have invented a novel evolutionary programming algorithm that can cut stocks efficiently by minimising the wastage [12]. The algorithm performed significantly better than other algorithms on a number of problems we have tested [12]. The algorithm can be used to cut any one dimensional stocks.
Container packing could be regarded as a 3-d version of stock cutting. The problem occurs daily in transportation and logistics companies. The key issue here is: How can we pack as many goods (of different sizes) as possible to large and fixed-size containers so that we use as few containers as possible? We have developed a software package on PC with a user-friendly GUI to perform container packing automatically. A user can enter the size of containers and sizes of the goods to be packed. With button-clicking, the software will display the packing that it generates using 3-d graphics.
Designing new materials, such as alloys, with certain properties is always a difficult and tedious task because we do not know all the properties of the material until we physically produce and test it. Material modelling is extremely important in understand various properties of the material. We have teamed up with a material scientist and developed a novel evolutionary approach to material modelling [13,14]. Our approach has been able to obtain the results that no other methods could. We have also developed a software package for material modelling. The software is generic and not limited to any particular type of materials.
Evolutionary algorithms are very good at discovering novel designs that can hardly be hit upon by a human designer. We have invented (UK and EU patent pending) a new method based on multiple Pareto fronts to evolve various designs [15,16]. The advantages of our method have been illustrated by a novel digital filter that we evolved [15,16]. It is worth noting that our method is not limited to digital filter design. It can be applied to other design tasks.
Finding the shortest as well as smoothest route between two locations is extremely important in designing new roads or pipelines for gas, water and oil. The problem is often complicated by the constraints, such as natural reserves, residential areas, archaeological sites, etc. We have developed an evolutionary algorithm based system for finding the shortest as well as smoothest route between any two locations in a map [17]. Comparisons with human designers have shown that our evolutionary system can find better solutions than human beings. The system can be used for many design as well as planning tasks.
Generating a timetable for a university or a sport event (e.g., the Olympic Games) can be a nightmare for the administrators because the number of courses/events and constraints involved. We have developed a timetabling system for the Australian Defence Force Academy [18], which produced results that were better than those generated by the timetabling secretary. Our system is also flexible and can be adapted to other requirements from a different organisation, e.g., another university or sporting event.
Computational Intelligence (CI) finds strong applications in manufacturing. CI techniques help us solve complicated problems for which no satisfactory solutions exist. These applications include:
A few examples of computational intelligence in manufacturering are presented here. More information is available.
Real-time monitoring of modern CNC machining equipment is hard but also essential to prevent damage to both machine tools and expensive work pieces. We have used a combination of Acoustic emission (AE), Cutting Force, Motor Current and Vibration Sensors to monitor real-time performance and accuracy of machine tools. We use a variety of CI techniques to analyse the data captured by these sensors and real-time control based on this information is able to improve the accuracy of the controlled machine considerably without a significant increase in investment.
X-ray based inspection systems are widely used for the identification and evaluation of internal defects, such as cracks, porosities and foreign inclusions in castings. We have used Wavelet Transforms and Imaging Processing techniques to detect internal defects using X-ray images and an improvement over traditional techniques.
Sheet metal stamping is an extremely popular manufacturing process. In-process detection of the malfunctions and on-line monitoring can help ensure product quality and also help protect expensive dies and presses. This is particularly important for unmanned modern stamping operations.
