Dr David Toll Geotechnical Systems Group, University of Durham, UK

Knowledge-based systems and neural networks are introduced within the context of geotechnical engineering applications. The knowledge-based systems that have been developed in geotechnical engineering (since their first reported use in 1985) are categorised according to their areas of application. More recent developments in neural netowork applications are also included in the analysis. The paper then reviews knowledge-based systems (and one neural network system) that have been developed for ground improvement and geosynthetic applications. It is concluded that systems developed using artificial intelligence techniques should aim to fulfill the role of decision support tools or 'assistants'. Hybrid systems which combine different artificial intelligence techniques and traditional algorithmic approachs will probably best serve that role in geotechnical engineering.

Knowledge-based systems (KBS) are suited to developing applications which make use of heuristics, or empirical knowledge, rather than solving a set of mathematical equations. They are oriented towards symbolic processing (words) as opposed to conventional programming techniques (which operate best with numbers). Another important feature of a KBS is that the 'control' aspect of a computer program (ie the computer code that controls what the program will do next) is a separate entity, called an inference mechanism. This ensures that there is a separation between the 'knowledge' and the inference mechanism. In conventional, procedural programming, data and 'control' are mixed together within the computer code. This separation allows knowledge to be added/modified/removed without having to change the computer code.

Rule-based approaches provide the most straightforward form of reasoning that can be used in knowledge-based systems. IF ... THEN ... rules are used. While each individual rule may in itself be very simple, more complex reasoning is achieved by chaining rules together. Different reasoning strategies can be used. Forward chaining reasons from known facts and rules, following all possibilities to see what conclusions can be reached (inductive logic). Backward chaining starts from a given goal and searches for facts or rules to support it (deductive logic). Mixed chaining can also be used which combines forward and backward chaining.

Case-based reasoning is an alternative to rule-based reasoning. Here, the new problem to be solved is matched with existing cases in the knowledge base. Cases stored in the knowledge base are 'indexed' so they can be recalled using search techniques. Similar cases can therefore be recalled, and decisions for the new problem can be based on what has been done before. Abstraction techniques can be used to 'adjust' cases when the match is not identical.

Semantic nets can also be used. This involves representing knowledge within an explicit structure, thus giving it 'meaning'. The concept of structuring knowledge can be used for facts or rules. Object oriented programming provides a convenient means of implementing either of these forms of knowledge representation.

A geotechnical example of structuring facts is shown in Fig. 1 in which soil types are structured within a hierarchy (Moula & Toll, 1993). The explicit structure of the facts within the knowledge base allows a system to reason about a soil description. By matching a material described as 'clayey sand' against it's knowledge base it can deduce that it is a form of 'sand', that it is 'coarse' and 'inorganic' in nature, and is a 'soil'. Properties of each soil type can also be stored within the knowledge base. This form of knowledge-base has been used within a KBS for estimating geotechnical properties (Toll & Giolas, 1995) and also forms part of a larger system for interpreting site investigation information (Toll et al, 1992; Toll, 1995).

Fig. 1. An example of structuring facts - a soil type tree

Rules can also be structured, using rule frames to group them together. A geotechnical example is given by Toll & Barr (1995). This involves a KBS (called ConFound) for conceptual design of foundations. It has been developed as computer-aided learning tool and is intended to assist undergraduate students in learning about the process of selecting a foundation type appropriate to the ground conditions. In this example, the decision making process that a student should follow is represented as a hierarchy, as can be seen in the left-hand side of the screen shot in Fig. 2. This hierarchy is explicitly represented within the system's knowledge base. The rules that identify the suitability of each particular foundation type are stored as a rule frame within the object representing that type.

Fig. 2. Screen shot of ConFound - a KBS for conceptual design of foundations

Uncertainty is frequently an issue within knowledge-based systems. Simple rule-based systems may only use first order logic (a deterministic aproach) in which outcomes are either 'True' or 'False' (no uncertainty). Uncertainty may still be present in the system (in an implicit form) due to the use of qualitative terms which are not precisely defined. However, such a system cannot provide quantitative information as to how reliable any conclusion is.

Knowledge-based systems can incorporate uncertainty in a variety of ways. Fuzzy logic can be used in which outcomes have 'certainty factors' or probabilities attached to them. Common approaches make use of Bayes' theorem of probability, or the more complex Dempster-Shafer method. Zadeh's Fuzzy Set approach has also gained a strong following among KBS developers. However, there are frequently difficulties in identifying what values of certainty factors or probabilities should to be used within a system. Experts are often unable to quantify the ways in which they handle uncertainty. Therefore, developing 'uncertain' systems involves a considerably greater effort in knowledge acquisition, and such systems are more difficult to validate.

Neural Networks are another form of artificial intelligence system that work best with quantitative data. Using a neural network involves developing relationships between data by 'training' the network on existing cases. Input data and known outcomes are used in training. Once these relationships are established they can be used to predict an outcome for a new case. Once the outcome is known then the network can 'learn' from this new case, and update the implicit relationship within it.

A review of knowledge-based system applications was presented by Moula et al (1995). This has been updated and extended by Toll (1996). The first geotechnical applications of knowledge-based system technology date from 1985/6. However, over the last decade a significant number (103) of knowledge-based applications have been developed. Neural networks were first applied in geotechnical engineering in 1991/2, but in the past two years that there has been a big increase in the number of neural network applications being reported (28).

 Fig. 3. Artificial intelligence applications in geotechnical engineering, classified according to the area of application.

The largest number of systems has been developed for foundations (25 in all, of which 4 are neural networks). However, this group represents quite different types of application including conceptual design, detailed design, construction methods, construction on problem soils and pile driving. Therefore, if one looks in more detail at the particular focus of the systems, then site investigation interpretation has attracted the most development, with underground openings in rock and conceptual design of foundations also representing significant areas of development. Neural networks have been developed particularly for classification and parameter assessment in both soils and rocks, for design of underground openings in rocks and for pile driving.

To reflect the theme of the conference, the following is a brief review of the artificial intelligence systems developed for ground improvement, including the use of geosynthetics.

IMPROVE (Chameau and Santamarina, 1989). This is a KBS designed to assist in the selection of soil improvement techniques. It is intended to help the user decide if there is need for soil improvement and to help select the best soil improvement technique. Although it attempts to find the 'best' solution it can continue the search for less satisfactory solutions at the user's request. It also uses a case-based system to select case histories that best resemble the project (50 case histories) and provides a postprocessor to provide final information and suggestions.

ESPGIS (Motamed et al, 1991). ESPGIS stands for 'Expert' System for Preliminary Ground Improvement Selection. It can advise on the selection of ground improvement methods or it can evaluate the suitability of a user's preselected method. It allows the user to define (with varying degrees of certainty) the nature of the ground improvement needed, subsurface conditions and other relevant parameters. It is also capable of assigning typical values for design parameters to soils based on soil descriptions and index properties.

Knowledge database for Ground Improvement (Yoon et al, 1994). This is described by the authors as a 'knowledge database' for ground improvement technologies. Strictly it is a database as there is no inference mechanism. However, it is rich in knowledge and could be developed into a KBS. It contains information on the current technologies available, classified by country of use and application. The information relates to: international/national codes of practice; design methods; state of practice; case studies.

PACT (Kotdawala & Hossain, 1994). This is a knowledge-based system for soil compaction that has been developed to identify the lift thickness and moulding moisture content for field compaction. It has knowledge of the different types of compaction plant as well as knowledge of problems associated with compaction of particular soils.

Neural-Network For Soil Compaction (Basheer & Najjar, 1995; Najjar et al, 1996). This is a neural network approach for soil compaction. It is intended to be used for predicting optimum moisture content (OMC) and maximum dry density (MDD). The input data to the neural network is soil type, grading characteristics and consistency limits. For natural soils prediction is based on only three input variables: liquid and plastic limit and specific gravity

Knowledge-based System for Geosynthetics (Maher and Williams, 1991). This is a KBS for selection and design of Geosynthetics. It assists the user with selecting an appropriate geosynthetic material and can perform detailed designs for different geotechnical applications. It contains knowledge about material selection for: stabilisation to reduce erosion; separation of soil layers; reinforcement; drainage; filtration to reduce cross plane flow of soil particles.

EDxES (Dimmick, Bhatia & Hassett, 1991). EDxES stands for Edge Drain design and specification by 'Expert' System. It can assist in the design and specification of the geotextile component of an edge drain for a road pavement. It considers commercially available non-woven geotextiles that perform the dual functions of drainage and separation. The output consists of the required hydraulic and mechanical properties and a list of the ten thinnest (lightest) candidate products.

Knowledge-based System for the Design of Geotextiles (Mannsbart & Resl, 1993). This is a KBS for design of Polyfelt geotextiles, based on the design charts in Polyfelt TS Design and Practice. The areas of application include: road construction; hydraulic construction; drainage systems; retaining walls; geomembrane protection. It can provide background information on: geotextile functions; geotextile properties; specifications; useful construction hints.

ROAD (Dukes et al, 1994). ROAD is a KBS for geotextile based road design. It covers design of primary and major road highways, based on AASHTO design procedures. It allows the inclusion of a geotextile layer and considers both the mechanical and filtration properties of the geotextile.

In all, 131 artificial intelligence systems have been developed with applications in geotechnical engineering. These are predominantly knowledge-based systems (103), but the number of neural network systems has increased rapidly since 1994 so that there are now 28 geotechnical applications. Many of these systems are simple developmental prototypes, with limited amounts of knowledge. They would need considerably more development work before they could be used in practice. However, some systems are now going beyond that early stage, and we could see useful AI systems available in geotechnical practice before too long.

The pioneers of knowledge-based systems (the computer scientists) saw the development of 'expert systems' (as they were then generally known) as an attempt to replace human expertise with computer systems. As developments proceed is this field, those involved are becoming more realistic about what can be achieved with the technology. AI systems can provide useful 'assistants' to engineers to help them with their tasks. Therefore, the most useful role they can play is to perform the more routine tasks, thereby leaving the engineer with more time to devote to the more difficult (and usually more interesting) problems.

AI techniques are very good at solving problems based on heuristic (empirical) knowledge. However, geotechnical engineering frequently requires a combination of heuristic reasoning and calculation. Many systems that have already been developed are hybrids, combining a knowledge-based component with calculation routines using traditional algorithmic approaches. This type of development is probably the best scenario for engineering systems, and we are likely to see more hybrid systems. These systems will not only combine AI and calculations, but will also use a mix of AI techniques.

A significant number of artificial intelligence (AI) systems have been developed for geotechnical applications. These are predominantly knowledge-based systems but the number of neural network systems is increasing. These include a small number of systems for ground improvement/geosynthetic applications.

Some geotechnical AI systems are progressing beyond the developmental prototype phase, and we should soon see AI systems in geotechnical practice. AI systems should be developed as decision support tools or 'assistants', rather than attempting to replace human expertise. Hybrid systems that combine different AI techniques and algorithmic approaches are likely to the best solution for tackling engineering problems.