Dynamic Service Generation: Agent interactions for service exchange on the Grid. (Génération dynamique de service : Interactions entre agents pour l'échange de service Grid)

ion at the control level consists in defining new functions or procedures abstracting on theexisting ones. For example, in Scheme, expression such as: (define square (lambda (x) (* x x)))(define foo (lambda ...)) FIRST-CLASS CONSTRUCTS Scheme provides in particular a powerful and simple mem-ory model via first-class environments as well as a very flexi-ble and dynamic control model via first-class procedures andfirst-class continuations. The notion of first-class procedure isvery important because it allows program (i.e., complex pro-cedures) to dynamically generate other programs. first-classobjects can be named by variables, can be passed as argumentto procedures, can be returned as results of procedures andcan be included in (and retrieved from) data structures.a Forexample: (define multiply-by-n(lambda (n) (lambda (x) (* n x))))(define multiply-by-5 (multiply-by-n 5))(multiply-by-5 3) ⇒ 15 Here, the multiply-by-5 function is dynamically con-structed by calling multiply-by-n. This mechanism canbe implemented with other languages such as Java but it isvery heavy due to the fact that it is not intrinsic to the lan-guage.b The first-class view of procedures enables the pro-grammer to consider the delivery of a procedure similar to thedelivery of a datum.The term ’higher-order’ may also be sometime used (especially for function),but it does not include the first and fourth properties.In Java, objects are considered to be first-class but procedures/methods arenot (e.g., they can not be returned by other procedures or passed as arguments).However, the abstraction toolcan usually reach only the firsttwo levels. In order to abstract atthe interpreter level, what we callmeta-level abstraction or meta-level learning, the interpreter it-self has to be modified. Thisfeature is necessary in order, forinstance, to add some specialforms to the language (e.g., inScheme, an applicative language,the special form if must be im-plemented in the interpreter itself(i.e., at the interpreter level), itcan not be defined at the controllevel). Changing the interpreterlevel must be done as well, in or-der to change the way of inter-preting expressions, for example,in order to shift from an applica-tive order evaluation to a normalorder evaluation, or to a lazy-evaluation (section 1.4.3.5).The two first levels can eas-ily be reached dynamically dur-ing execution: new data and pro-cedures are added to the languageeach time an abstraction is in-terpreted or an assignment done.The challenge is to allow inter-preter level modifications at runtime, in order to generate dy-namic processes. When the interpreter level is dynamically changed, the rest of the computation isinterpreted with the new interpreter.Actually, we may say that these three levels help us to define a language as a environment-interpreterpair. The interpreter is the program or mechanism that gives a sense to an expression (evaluates) in aparticular context. An expression of the language is an expression recognized by the interpreter. Thiscontext is called an environment; it is used as a memory to store abstractions constructed with the lan-guage. Therefore, to enrich the language means being able to abstract (learn) at the three cited levels. TheSTROBE model maps these three levels schema into agents, considering that they are composed of a setof pairs of cognitive environments and cognitive interpreters. Data and control levels learning are madeby changing the cognitive environment and meta-level learning by changing the cognitive interpreter.