Get your own Free Home PageInductive learning techniques (ILT's) have grown to technical maturity over the last two decades. Many inductive learning techniques (e.g. decision tree learners, neural networks, case based reasoners, ..) are now off-the-shelf available, performing well on learning tasks. The functionality of most of these techniques covers classification and regression. From other disciplines, such as statistics, techniques with similar functionality are available.
In application projects where ILT's have been used, problems occur on three levels: identifying promising opportunities for ILT application , mapping of tasks on techniques and tuning of techniques. Recently we have studied the ILT application practice in the Netherlands. Based on this work a proper understanding of application problems has emerged, leading to the (up to now) tentative formulation of a process model, the Method for Designing Inductive Applications (MEDIA). Additional work has been done to facilitate this process model with dedicated knowledge on the functionality of learning techniques.
In 1997 I was co-organiser, together with Robert Engels, Bob Evans and Juergen Herrmann, of the workshop Machine Learning Application in the Real World: Methodological Aspects and Implications, hosted by ICML'97, in Nashville (TN).
A follow-up workshop has been announced, to be held in 1998 at AAAI and ICML (Madison Wisconson). This workshop entitled The Methodology of Applying Machine Learning will focus on three important subproblems in ML application: Problem Definition, Task Decomposition and Technique Selection. In this workshop, the other organisers are Robert Engels and David W. Aha.
In recent years I have worked on two applications in (semi)agricultural domains: planning of product treatment and resource allocation. The first system, the Product Treatment Support System (PTSS), supports human experts in planning treatment conditions for agricultural produce. We define treatment as the process of non-destructively controlling the product state by exposing external conditions (e.g. temperature, relative humidity, concentration of ethylene, CO2 and O2, etc). that are imposed on a product. Typical treatment processes are ripening of fruits, long term storage of products (e.g apples, pears, potatoes) and preparation (flower bulbs).
The system actually solves the problem in two steps: quality assessment and recipe assignment. In order to allow feasible quality measurement, experts have to cope with huge variations within batches. By providing them with simple rules for guiding the sampling process, generated by Quinlan's C4.5, sample size could be reduced from 50-200 to 5, making quality measurement feasible in practice. The recipe allocation is itself a two-step process. First, a neural network assesses total recipe requirement from current quality and required post-treatment result. Then, a constraint satisfaction module assigns conditions to time-slices. The original PTSS is delivered as a prototype, with a proof-of-principle character. Several follow-up projects are underway, amongst others for planning conditions for exotic fruits in the logistic chain.
The second system is a scheduler for allocating aeration resources in a waste water treatment plant. This system, part of the WaterCIME project, takes a process status and a prediction of short term waste water load as input, and produces a short term schedule for the super-chargers. To overcome moving horizon scheduling is performed in a five minute cycle.
The system applies a case based approach. The case base contains realised schedules, process states and operational results. New cases, consisting of process states and predicted load trigger retrieval of similar cases. Retrieval is facilitated by an indexing mechanism based on Self Organising Maps. This allows knowledge extensive indexing. The SOM can however be used to generate a posterior knowledge about the structure of the schedule space. The schedule of the retrieved case is heuristically adapted to fit the actual situation.
In 1996 I was initiator and co-organiser of the workshop Neural Network Applications in Agriculture (NNAA), hosted by the International Congress on Computer Technology in Agriculture, ICCTA'96, in Wageningen, (NL). Some of the papers of this workshop are currently under publication in AI Applications in natural resources, agriculture, and environmental science.
This page hosted by Get your own Free Home Page