Adaptation modelling?

The issue of modelling the behavior of any adaptive system has two complementary but distinct dimensions, which we will examine separately: the specification of adaptation logic, and the specification of  adaptation actions. The former is responsible for relating information available in one or more models and assessing whether adaptations are required. The latter refers to specifying the very actions that need to be effected by the system for a given adaptation to be achieved.

Attempting to standardise the way in which adaptation logic is expressed would be, in the authors’ opinion, rather premature at this point in time. Existing approaches include simple rule-based engines, case-based reasoners, etc., all the way to powerful logic-based reasoning engines. Given this wide range of approaches in use, it is apparently unrealistic to aim at a single specification that could accommodate them all. On the other hand, developing a range of specifications should be undertaken only after evolution in the targeted approaches has reached a critical level of stability, ensuring validity and endurance of the specifications over time.

Unlike the case of adaptation logic, adaptation actions constitute a well-researched and rather “crystallised” field, especially as far as Adaptive Hypermedia Learning Systems are concerned (Brusilovsky, 2001). Furthermore, recent research (Paramythis and Stephanidis, 2004) has proven the feasibility of formalising and declaratively specifying (using an XML-based language) adaptation actions to be effected as part of an adaptation cycle. It is argued that such efforts could easily be extended, so as to arrive at a standard that allows for flexibility as far as adaptation logic in concerned, and defines a concrete way for coupling that logic with an extensible set of adaptation performatives for ALEs.

Of the existing standards, the only one that supports the explicit representation of dynamic behaviour on behalf of the system is the IMS Learning Design (LD) specification. In more detail, Levels B and C of the specification under discussion introduce the concepts of properties, conditions and notifications, which can be used to specify arbitrarily complex dynamic behaviours for a system. The main setbacks in employing the IMS LD for modelling adaptation in ALEs are rooted in the fact that specification of dynamic behaviour is achieved through the definition of programming flows (including condition variables), enriched with event semantics:

􀂃 The approach can be considered rather low-level: Specifying complex adaptive behaviours is tedious and error-prone.

􀂃 Conditionals may only refer to variables or states that exist in the context of a single IMS LD document (which makes it impossible to consult models external to the document).

􀂃 Dynamic behaviors cannot be defined at the system level (and applied in more than one contexts, or for more than one sets of learning materials / activities).

􀂃 The dynamic behavior specified cannot be reused: there is tight coupling between the behaviour itself and the artifacts to which it refers.

􀂃 And, finally, the behavior specification lacks semantic-level information which would allow an ALE to modify or affect it in any way.

Despite the above shortcomings, the IMS LD may be a very appropriate vehicle for introducing adaptive capabilities in non-adaptive e-Learning systems. Specifically, an adaptation engine can be introduced in an LD-capable system, which would effect adaptations by generating or augmenting LD specifications “on the fly”. In other words, such an engine would translate adaptation logic and actions into IMS LD compliant constructs, which would then be delivered to the user. By going through this process dynamically (at run-time), the system would also be able to incorporate into the generated constructs, current information derived from adaptation-specific models.