Learning methodology

Sheikh, A.Y. and Jones, A.G., 1997. Crystallization process optimisation through a revised machine learning methodology. American Institution of Chemical Engineers Journal, 43, 1448-1457. 

Case-based reasoning is very much dependent on the structure and content of its cases and their representation because case retrieval involves identifying those features in the problem that best match those in the case base. The dynamic addition of new cases means that CBR is intrinsically a learning methodology such that the performance of an expert system based on this approach will improve with time [9]. Systems may be developed with conventional computer languages or shells [7]. 

Thus, a critical departure from previous approaches, common to all our learning methodologies, is the adoption of a solution format that consists of hyperrectangles (not points) defined in the decision space. 

The notation introduced above allows us to make now a more explicit and condensed enumeration of the major characteristics and differences, with respect to the Q, ij/,f,S) key components, that separate our learning methodologies from other approaches. 

As we will see in subsequent sections, the mapping procedures, /, adopted in our learning methodologies, are based on direct sampling approaches ... 

A solution space, a, consisting of hyperrectangles defined in the decision space, X, is a basic characteristic common to all the learning methodologies that will be described in subsequent sections. The same does not happen with the specific performance criteria tfi, mapping models /, and search procedures 5, which obviously depend on the particular nature of the systems under analysis, and the type of the corresponding performance metric, y. 

In this section we will introduce the problem statements adopted for this type of performance metric, briefly describe the learning methodology employed to address it [for a more complete presentation, see Saraiva and Stephanopoulos (1992a)], and show a specific application case study. 

To illustrate the potential practical capabilities of the learning methodology, we will now present the results obtained through its application to... 

This Section addresses cases with a continuous performance metric, y. We identify the corresponding problem statements and results, which are compared with conventional formulations and solutions. Then Taguchi loss functions are introduced as quality cost models that allow one to express a quality-related y on a continuous basis. Next we present the learning methodology used to solve the alternative problem statements and uncover a set of final solutions. The section ends with an application case study. 

Consequently, the goal of our learning methodology is the identification of hyperrectangles in the decision space, X, that minimize expected total manufacturing cost, (y X), a performance measure that combines in a consistent form and a quantitative basis both operating and quality costs. 

In the previous paragraphs we defined the solution format f, performance criterion i/r, mapping procedure /, and performance metric y that characterize our learning methodology for systems with a quantitative metric y. Here we will assemble all these pieces together and briefly discuss the search procedure, S (further details can be found in Saraiva... 

In order to verify how close to a known true optimum the final solutions found by our learning methodology happen to be, we will describe here its application to a pulp digester, for which a perfect empirical model /(x) is assumed to be available. Other applications are discussed in Saraiva and Stephanopoulos (1992c). 

To support the application of the learning methodology, fix) was used to generate 500 (x, 2, w) records of simulated operational data, transformed by Eq. (26) into an equivalent number of (x, y) pairs. Finally, the following constraints were imposed to the search procedure, 5 ... 

To benchmark our learning methodology with alternative conventional approaches, we used the same 500 (x, y) data records and followed the usual regression analysis steps (including stepwise variable selection, examination of residuals, and variable transformations) to find an approximate empirical model, / (x), with a coefficient of determination = 0.79. This model is given by... 

In this section we describe extensions of the basic learning methodologies introduced in Sections IV and V that, while preserving the same premises and paradigms, enlarge considerably their scope by adding the capability to consider simultaneously multiple objectives. As before, and without loss of generality, we will focus our attention on the coexistence of several quality-related objectives. 

Both situations with categorical and continuous, real-valued performance metrics will be considered and analyzed. Since Taguchi loss functions provide quality cost models that allow the different objectives to be expressed on a commensurate basis, for continuous performance variables only minor modifications in the problem definition of the approach presented in Section V are needed. On the other hand, if categorical variables are chosen to characterize the system s multiple performance metrics, important modifications and additional components have to be incorporated into the basic learning methodology described in Section IV. 

Since these loss functions express quality costs on a common and commensurate basis, extending the learning methodology of Section V to a situation with P objectives is straightforward. All one has to do is replace the original definition of the y performance metric [Eq. (23)] by the following more general version ... 

The aspiration levels inherited from the previous step, y, are used to guide the search process. Each agent i employs the corresponding aspiration level, y, and through the application of the learning methodology presented in Section IV tries to identift feasible hyperrectangles, Xf, that lead to performance consistent with y. ... 

To conclude this section on systems with multiple objectives, we will consider a specific plasma etching unit case study. This unit will be analyzed considering both categorical and continuous performance measurement variables. Provided that similar preference structures are expressed in both instances, we will see that the two approaches lead to similar final answers. Additional applications of the learning methodologies to multiobjective systems can be found in Saraiva and Stephanopoulos (1992b, c). 

Complex manufacturing systems, such as an unbleached Kraft pulp plant (Fig. 9), are almost always characterized by some type of internal structure, composed of a number of interconnected subsystems with their own data collection and decisionmaking responsibilities. This raises a number of additional issues, not addressed in previous sections. For instance, if the learning methodology described in Section VI is applied to the digester module of a pulp plant (Fig. 9), it is possible for the final selected solution, to include ranges of desired values of sulfidity... 

To address the modified problem statements and uncover final solutions with the desired alternative formats, data-driven nonparametric learning methodologies, based on direct sampling approaches, were described. They require far fewer assumptions and a priori decisions on the part of the user than most conventional techniques. These practical frameworks for extracting knowledge from operating data present the final uncovered solutions to the decisionmaker in formats that are both easy to understand and implement. 

We presented extensions and variations of the basic learning methodologies aimed at enlarging their flexibility and cover a number of different situations, including systems where performance is evaluated by categorical or continuous variables, with single or multiple objectives, simple or complex plants containing some type of internal structure and composed of a number of interconnected subsystems. 

The potential practical capabilities of the described learning methodologies, and their attractive implementational features from an industrial point of view, were illustrated through the presentation of a series of case studies with both real-world industrial and simulated operating data. 

The next section will focus on the representation necessary to express this sufficient theory to the computer, so that it can automatically carry out the reasoning associated with analyzing the examples selected by the syntactic criteria presented in this section. Section V will describe the learning methodology, which, using the representation of Section IV, will generate the new dominance and equivalence conditions. 

It can be shown that the unsupervised learning methodology based on Kohonen self-organizing maps algorithm can be effectively used for differentiation between various receptor-specific groups of GPCR ligands. The method is similar to that described in Section 12.2.6. 

---