Contact: Pierre Dillenbourg
This project results from the congruence between the type of software developed by TECFA and the skills taught by FSS. In short, the characterization of petroleum reservoir requires a set of a complex skills whose acquisition heavily rely on experience. Therefore, a learning environment simulates a situation in which the trainee has to perform the same cognitive processes as in his future job. The best example of learning environment is a flight simulator. A learning environment is referred to as 'intelligent' when it includes an agent which can guide and advise the subject. This agent is generally implemented with artificial intelligence techniques.
Through this project FSS aims to maintain a training offer which is up-to-date both with respect to the content and the method. The ILE - called GEOLAB - will not substitute to courses, but enrich these courses and extend training beyond short sessions. This time extension is especially relevant because such complex skills cannot be learned in one week. Therefore, a key issue addressed in this project is the integration of the ILE with the existing lectures provided by FSS. We address this issue in two ways. On one hand, GEOLAB will contain and refer to the multimedia material presented during the lectures. For instance, when the user does some mistake, the system may show relevant parts of theory or analogous cases. On the other hand, GEOLAB will be articulated with the geostatistics software "FSS TOOLS" used by FSS for its training courses.
Over the last seven years, FSS International has established itself as a leader in the field of technical training and continuing education for the environmental and mining industry. For the oil industry alone, it has taught to over one thousand production geologists, petrophysicists, seismic interpreters and reservoir engineers how to apply geostatistical methods for analysing their data and developing numerical reservoir models. All the major oil companies (Shell, Mobil, Exxon, Esso, Amoco, Elf, Texaco, ...) have asked FSS International to provide in-house training, or have sent the professionals to FSS International public short courses. FSS offices are located in Calgary, Reno, Paris, Geneva, Vancouver and Sidney. For the project, the partner is FSS Consultants SA, the Swiss member of FSS International, directed by Dr. Roland Froidevaux.
Through the years, FSS International has endeavoured to constantly improve and adapt the courses in order to be responsive to the practical challenges of reservoir characterisation and to ensure that the training continues to present current improvement in methodology. This concern for innovation justifies FSS investment in this project. Another source of motivation for FSS is the possibility to extend training beyond short sessions. Currently, the longest training sessions delivered by FSS last five days, including lectures and workshops. Despite the fact that these sessions are very intensive, trainees would benefit from longer working sessions. GEOLAB will enable participants to pursue their training after the end of the training session. This continued education will not only increase the efficiency of the training provided by FSS, but also enable FSS to maintain a longer relationship with their customers.
TECFA is an multidisciplinary research unit in educational technology. Our motivation for this project is two-folded. One on hand, we want to transfer to real scale problems the techniques and prototypes developed in previous research projects. We acquired expertise in the field of 'intelligent learning environments', i.e. advanced educational software using artificial intelligence techniques. During this research, we paid attention to the generality of algorithms in order to support portability across domains. It is hence now our primary interest to test our previous work in a different field. The second interesting aspect for TECFA is the issue of integrating the ILE to be developed with existing lectures and workshops. Too often, designers have built systems without considering how they would be integrated in the existing courses. This lack of concern for integrating new software within existing activities has been detrimental to the development of educational technology.
Moreover, this project would lead TECFA members to play an innovative role in the Swiss community of training software ("Schweiz CBT"). TECFA belongs to this association and has published several papers in the association journal. Most training software we observed is 'superficially' interactive: designers include nice sounds and images, but do not pay enough attention to the cognitive activities of users. Nobody learns complex skills simply by answering questions. We hope that GEOLAB will convince developers to build richer interactive systems, or at least inform the companies who buy such systems that richer forms of interactivity can be developed.
To illustrate this problem, let consider a typical dome shape reservoir (Figure 1). The total volume of this reservoir (which is the single most critical parameter for the reserve estimation) is given by the intersection of its top surface (in general corresponding to some stratigraphic or structural horizon) with an horizontal plane corresponding to the oil-water contact. These two surfaces (oil-water contact and top of structure) are known precisely at well location only. Elsewhere, they will have to be estimated, partly by extrapolating from the well locations, and partly by interpreting available seismic information. This estimation process involves a whole series of assumptions (on the spatial correlation, wave velocities, rock densities, etc...) and each of them is based on expert judgement, i.e. subjective.
As a consequence both surfaces are uncertain as to there exact location. Figure 1 illustrates two scenarios for each measure: two variations of the oil-water contact (OWC+ and OWC-) and two variations of the top of structure (TOS+ and TOS-). The impact of this uncertainty is severe on the resulting total volume: the volume corresponding to the optimistic case (light grey area) is in the order of three time bigger than the volume corresponding to the pessimistic case (dark grey area).In the decision-making process (should the oil field be developed or not), the manager will have to consider not only the most likely scenario, but also (and possibly especially) the pessimistic one. If the economics of this low case scenario are unfavourable enough, he may decide not to invest in the development of the reservoir, although the estimated reserves (most likely scenario) were promising. Uncertainty characterization, therefore is a crucial element in reservoir management and an improper assessment of it may lead to huge financial losses, either by being too pessimistic (rejection of a good prospect) and foregoing revenues, or by being to optimistic and sinking capital cost in a project that will never be economic.
It is clear therefore that the characterization uncertainty is not a trivial task which can be learned on the fly. And since it is not a topic normally found in academic curricula for earth sciences, oil companies rely more and more on in-house professional training to make sure that their key personnel has the required level of familiarity with probabilistic and geostatistical concepts necessary for taking decision in face of uncertainty. Training professionals at characterizing uncertainty fits the approach and systems developed at TECFA for the acquisition of complex problem solving skill. The adjective 'complex' refers to the four following aspects of the task described above:
Uncertainty characterization requires familiarity with a series of concepts and techniques, which form the framework of applied geostatistics. Among the key concepts are the notions of spatial correlation, of prior and posterior distributions (Bayesian theory), of loss functions, of fuzzy variables, etc... The tools offered by geostatistics, and which are based on these concepts, can be subdivided into two broad categories: estimation tools and simulation tools, each category being further subdivided between parametric or non-parametric techniques. These concepts and techniques are currently taught at FSS International short courses, and will be progressively integrated, through problem solving in GEOLAB.
The learner must maintain simultaneously in working memory all the elements necessary to solve the problem. The GEOLAB system will help to reduce the cognitive load in two ways. On one hand, the computation will not be performed by the learner, but by FSS TOOLS. On the other hand, the learner will use a 'progress pad' in which she will note intermediate results or decisions.
Characterizing uncertainty does not reduce to a single 'cookbook recipe'. It relies on heuristic knowledge to select the appropriate technique given the specifics of the problem at hand. For instance the problem of 'merging' direct and indirect information (well data and seismic data) may be addressed by half a dozen of techniques at least. The selection of the appropriate one will depends on the amount of direct information available, the nature of the seismic information (whether it is a high or low frequency attribute) and whether some other external constraints should be applied or not. One of the critical issue is the amount of 'expertise' injected in the process. As the amount of direct (conditioning) information decreases, one has to rely more and more on a-priori assumptions from the experts.
The challenge of uncertainty characterization is to come up with a probabilistic description of potential errors. By definition, this distribution can never be verified experimentally. If one could live long enough, one could hope to know what is the 'true' volume of the reservoir. And since one knows also what his estimated value was, one can establish what the error is. But one cannot build a distribution of errors from a single one! Because of the impossibility to check a-posteriori the assessment of uncertainty, the only way to exercise quality control is to be critical on the assumptions made and on the methodology adopted.
Each category of software has its defenders and its disclaimers. It is pointless to argue that a category is better than another. The real question is to assess whether the learning activities of some program correspond to the objectives fixed by the designer. We argue here that intelligent learning environments are best suited to teach such complex problem solving skills such as those selected for this project (Dillenbourg et al., 1994b).
Jumbo pilots are not trained (only) with books, nor by answering simple questions. Traditional educational software, based on questions and answers, is suited to the acquisition of declarative knowledge, for instance the basic terminology of the field. One can teach simple skills, like computing a mean, with a simple 'drill and practice' courseware. But, for complex skills such as conducting an aircraft or reservoir characterization, it is not enough to know concepts or procedures. The trainee has to solve step-by-step the problems she will face during her career. She must be allowed to explore a problem, try some solutions, observe the consequences of her actions, and do it again until she gets the right solution. An ILE provides the learner with experience and experience is the best way to acquire complex problem solving skills .
Jumbo pilots are not trained on helicopters! Learners often fails to transfer to real life the skills taught in courseware. This difficulty reduces dramatically the system efficiency, since the ultimate goal is not that the learner simply masters the course (post-test scores), but it becomes better in her work (Dillenbourg et al, 1990). When a subject learns how to solve problems from a class of problems P, many teachers or designers expect him to transfer spontaneously his skills to P', a set of problems viewed as similar by the teacher. Often, this does not work, because the difference between P and P' involves some implicit knowledge that the designer may not be aware of, but which prevents the learner to reuse the acquired knowledge. Jumbo Pilots of Airbus 320 are trained on an Airbus 320 and not on another plane, even similar. In the case of flight simulators, the learning environment reproduces as much as possible the physical features of the real workplace. In our domain, we restrict ourselves to a cognitive fidelity: the environment has not to reproduce the exact features of the future trainee's workplace, but it must require the same cognitive processes.
If a learning environment is as complex as the real world, it may miss its educational purpose: the learner may simply be overloaded by the number of things (s)he must think or do simultaneously. Hence, a learning environment should offer the possibility to simplify temporarily the problem. For instance, the system WHY (Frederiksen and White, 1988) includes three different worlds. In the last world, learner have to solve electricity problems with the Ohm laws. But, in the first world , the learner may conduct simple qualitative reasoning (current or not), and then move to the second world reasoning is semi-quantitative (increase/decrease of resistance, ...). In other words, a learning environment must support a temporary and adaptive simplification of the real world.
Even a simplified world may reveal to be too complex for the learner. In complex problems such as those mentioned here, it may be poorly efficient to leave the learner trying a thousand different solutions. She may never come out with a solution if she does not receive the guidance from a more experienced partner. Future pilots use flight simulators under the supervision of experienced pilots who can explain what they did wrong and suggest a better approach. Even if learning complex skills is primarily based on individual experience, it cannot be accomplished without interaction with more experienced partners. A learning environment is called 'intelligent' when the software includes a computational agent (or 'expert') able to interact with the learner during problem solving: to make suggestions, criticisms, ... To conduct such interactions, the agent must himself be able to solve the problems submitted to the learner. These agents are implemented with artificial intelligence techniques, namely knowledge-based systems. However, the technology of expert systems is not fully appropriated to educational purposes. Most expert systems reason independently from the user, who simply provides the initial data and receives later the results with some explanation. TECFA has adapted these techniques to make the expert system more interactive (Dillenbourg et al, 1995).
FSS lectures heavily rely upon multi-media material, mostly fixed schema and pictures (slides), but also short video scenes (especially for case studies). We plan to create links between the expert's knowledge base and this material. We applied a similar approach in the MEMOLAB system described above. We then faced the issue to connect a learning environment with an hypertext including some theory about human memory and the methodology of research. The explanations provided by the expert in MEMOLAB include links to the hypertext, simply presented as buttons in the explanation text. If the learner selects the links, she obtains more detailed information on the particular point of the explanation. She is not supposed to browse all the hypertext, but to access more directly to some information she needs to solve the problem. We aim to apply to same idea to the system to be implemented. The expert rules will be connected to the material presented during the lectures. The various elements of this material will themselves be connected among each other in such a way that the learner can browse it around his or her entry point.
FSS short courses are structured around lectures and workshops. The latter are based on a software called FSS TOOLS, a widespread software for geostatitics developed by FSS. This system is not a training software, but a set of tools for spatial data analysis and modelling. They include exploratory data analysis and modelling estimation (kriging, cokriging, etc...) and simulation (multi gaussian simulation and indicator simulation). Since these tools are an inherent part of the problem solution, it is highly desirable to develop the ILE around this existing software. This has strong technical implications on the implementation of the ILE: the ILE will be programmed as a module to plug on FSS TOOLS, instead of as a closed and independent system. This corresponds to the evolution of research on ILEs. Most ILEs so far have been implemented as closed system, often in Lisp or Prolog. This was the case for MEMOLAB. This method is not optimal because Lisp and Prolog, even recent releases, are less powerful than other tools to build interactions with the user. The current evolution in ILEs is to restrict AI tools to the implementation of the expert module. This also fits with a global trend in micro-computing, to increase the connectivity between applications, through DLL's, XCMDs, Apple Events, and so forth.
Given these specifications, GEOLAB will be developed will different tools (in the Windows environment):
The scenario involves three agents, the learner, an experienced colleague who is expert in the domain and a manager who provides them with problems and ask for results. The expert and the manager will be implemented as knowledge-based systems, with CLIPS. The learner has two tools available, a 'calculator' which is an interface to FSS TOOLS and a 'progress pad' , in which the learner fill structured progress reports, store the data computed with FSS tools and add any additional comment. The tools will be implemented with Visual C++. The scenario is structured around four progress meetings. The structure of these meetings will be reflected in the 'progress pad' that the learner will fill in during problem solving. In each of these meeting, the manager and the learner interact about the state of the problem. The scenario starts by a first meeting where the problem is set.
The scenario is extensible to other agents. We envisage the possibility to have human experts from FSS to interact with learners after the session via electronic mail or more recent synchronous tools such as the MOOs. This component is however not part of this current project.
Recording protocols of experts doing reservoir characterization (alone and in pairs)
Interviewing experts doing reservoir characterization
Modelling the knowledge used by experts (rulebase)
Modelling the most frequent mistakes performed by learners, according to the teaching experience of FSS experts.
Building the interactive expert system with CLIPS
Validating the system behaviour with human experts.
Improving the multimedia material used in FSS training sessions
Increasing the multimedia material: adding video reports from oil exploitation area, recording videotapes for critical explanations provided in lectures.
Organising the multimedia into an hypertext accessible from the ILE and browsable by the learner (with MOSAIC)
Connecting the multimedia material to the expert knowledge (adding hypertext links to the rules)
Creating a database of problems.
Implementing the 'progress pad' (with Visual C++).
Implementing the 'calculator', i.e. the interface with FSS Tools (with Visual C++).
Implementing the manager which conducts progress meetings (with CLIPS).
Implementing the interface between the learner and the expert (with Visual C++).
Implementing the interface between the learner and the manager (with Visual C++).
Experimenting the first release of GEOLAB, with individual subjects
Modification of the system according to these first observations
Experimenting GEOLAB during FSS courses