Once experim ents have been designed, constructed and carried out, th e results m ust be observed. N aturally occurring phenom ena m ust also be observed, as in im plicit passive experim ents m entioned above. On first consideration, it seems th a t observation is simple and straightforw ard, requiring merely th a t the app ropriate events be recorded. This is naive. Paradoxically, observation is the simplest and th e m ost difficult of the stages in the framework. A t the naive level, observation simply involves w aiting and watching. However, there are two kinds of objection to this view which we can call technological
and theoretical objections.
Technological objections are im portant b ut straightforw ard. Observation of the world by a machine observer is possible, but still very lim ited. A lthough th e term observation implies ju st vision, observation in a broad scientific sense applies to all perceptual abilities, including vision, smell, sound, and so on. The problem of integrating these different capabilities, each at an adequate level of efficiency and accuracy is a separate and difficult research area of its own, but one which has an im p o rtan t bearing on the possibility for independent observation as p art of a complete architecture. Technological observations apply also to hum an observers, in the ability of the actu al perceptual organs, and there are grounds for questioning the observations of such observers. Hum ans are also subject to lim itations and deficiencies in their perceptual abilities. (W hat is or is not observable
is a related b ut difficult question [1 2], which will not be considered here.)
This is related to the second kind of objection, theoretical objections, so called because they refer to the role of the theory in the process of observation. Two observers of equal perceptual ability m ay provide different observation statem ents of the same phenomenon depending on their prior experience. M any exzunples are available where the im age seen by two people is the séime, but their interpretation of it is different (eg. [5]). It is argued th a t observation is possible only in the light of some theory, and th a t the observations will be expressed (and interpreted) according to th a t theory. O thers such as Hacking [34], however, claim th a t theory is not necessary for observation. We will n ot dwell on the issue of the theory dependence of observations, bu t recognise th a t there is a contentious issue here th a t m ay not be resolved, and which causes difficulties in any discussion of observation.
5 .4 .1 O b se rv a tio n in M ID
Like experim entation, observation is an external stage in th e six-stage framework. In MID, observation is modelled by the explicit provision by the user of observation state m ents to the system as input. (Observation cannot be entirely divorced from the eval uation stage, since they are to some extent interdependent, b u t the distinction is clear enough for them to be treated separately.) A fter the stages of prediction and experimen tatio n , the observations are provided and then com pared w ith the generated predictions.
Observational Data >
□ □
Predictions Observation Recorded Data Observationso o o
Figure 5.3: Observation in MID
Enter observations (end with return alone): > increase (heat std-object2)
>
Is the following prediction observed?: decrease (heat std-objectl)
Figure 5.4: Checking observations through prediction in MID
appropriately), however, MID will a ttem p t to re-observe by prom pting the user for the appropriate observations. This allows observation to be directed in the way th a t it is in real world scenarios, where expectation failures often ensure a m ore accurate check on observations. Moreover, it introduces th e notion of relevance^ sdbeit in a lim ited and incomplete way. A lthough the external observer is prim arily responsible for determ ining which events are relevant to the phenom enon under investigation, this dem ands th a t all predictions are regarded as relevant. Thus MID provides two kinds of observation: observation of naturally occurring phenom ena, and observation of controlled phenom ena. Figure 5.3 illustrates M ID ’s observation m echanism .
In the exam ple introduced earlier, two predictions were generated. Now, if only one of the corresponding events in the world is observed, say, then the user (observer) is prom pted for the missing observation. This is shown in Figure 5.4.
The observation stage also entails recording the details of the observations for subse quent use in providing a measure of the evidential support of theories, and in m aintaining historical consistency (see C hapter 8). In practically all other work on autom ated dis covery, observation is im plicit, and no distinct consideration of observation is typically given. This m ay not be a problem in the short te rm when the role of observation will be lim ited by technology. However, it fails to recognise th e distinct role of observation (or data-gathering) in the real world, and thus lacks completeness. In addition, it ignores the po ten tial of future development which m ay allow the integration of separate observation com ponents into other systems.
5.5
D iscu ssio n
Note th a t use of the scenario description for prediction implies the existence o f an ex perim ent at this initial point. P articu lar predicted effects are generated in the context of an experim ent. Thus the stages of prediction and experim entation are connected, and their ordering is not strict. In the im plem entation of M ID, the ordering m ight be differ-
eut, and prediction m ight be delayed un til the point of observation when a comparison w ith predictions needs to be made. The six-stage framework, however, in considering issues beyond this particular im plem entation, establishes a basis for this ordering. Con ceptually, predictions follow from a theory, and it is those predictions th a t are tested by experim entation. W hen im plicit passive experim ents take place, th e prediction (or expectation) exists before the experim ent is performed.
The consideration here given to the first three stages of th e framework has taken an external and em pirical viewpoint. It should be app aren t, however, th a t th e stages of prediction, experim entation and observation are ju s t as suited to systems where the stages are not external, nor explicit. In concept form ation, for exam ple, th e prediction th a t each new instance presented will be accounted for by th e current concept descriptions is im plicit. E xperim entation and observation involve th e generation and presentation of the instance to the system.
In this chapter, we have discussed the first three stages o f the six-stage framework in some detail, and described their instantiation in MID. P rediction and lim ited observation capabilities have been im plem ented, but experim entation has been om itted due to prob lems of excessive am ounts of knowledge th a t would be required for an effective system. This concept of experim entation is, however, lim ited. P r o v id in g directed and controlled design and construction of experiments restricts the possibilities substantially. By leav ing this stage open, m any options are available, including the observation of naturally occurring phenom ena and other alternatives which allow the acquisition of experim ental d ata from different sources.
A lthough the focus of the research described in this thesis is not centred around exper im entation, its relevance and significance is appreciated. Indeed, th e six-stage framework shows the relation of the experim entation stage to the rest of the inductive program m e. Work addressing the design and construction of experim ents is progressing, and the framework provides a basis for the integration of this an d subsequent work. Thus al though experim entation is n ot addressed in detail here, there is an awareness of the ‘bigger p ictu re’, smd the current work has been undertsdcen w ith th a t in mind.