Formalização e implementação de um mecanismo de interferência do paradigma orientado a notificações baseado em multiplicação de matrizes esparsas

Abstract

Rule-Based Systems (RBS) are computational systems that organize knowledge into fact-base elements and rules, in which changes in fact-base elements cause the approval/disapproval of different rules throughout the execution flow of a program. RBSs are typically implemented by traversing (computationally expensive) data structures to check whether fact-base elements have changed and to know which rules should be approved or disapproved based on these changes, causing temporal (i.e., repetition of the calculation of boolean expressions) and structural (i.e., repetition of code) redundancies. In this sense, the Notification Oriented Paradigm (NOP) proposes the division of the program into “intelligent” notifying entities, which notify their dependencies when their internal states change, eliminating searches with temporal and structural redundancies for rule inference. In short, when its internal state changes, an element of the fact base will notify only the rules that are relevant to it, that is, the rules that depend on it, to be approved/activated or disapproved/deactivated. The purpose of this work is to formalize, through algebra, the dependency between rules and the elements of the fact base of a NOP system as the multiplication of sparse matrices by dense vectors, or SpMV (Sparse MatrixVector Multiplication). From the formalization, a new NOP framework was developed in C++, called MBRIS 0.1 (Matrix-Based Rule Inference System), which implements a new notification-oriented inference mechanism, inspired by matrix multiplication to animate notifications. This is to the detriment of the usual way of implementing in which an entity has the address of the notifiable entities, the NOP, as in the NOP Frameworks that, at the time, allowed usual languages (e.g., C++, Java and C#) to work in a notification-oriented manner. In addition to the innovation in the notification method, aiming to take advantage of the graphics card hardware, specialized in performing operations with matrices in a highly parallelizable way, an OpenCL kernel was implemented in MBRIS to be able to execute the NOP notification flow on the graphics card. Finally, a sensor simulation application was implemented as a proof of concept using MBRIS, and the performance of the application in MBRIS was compared with the same application developed using the Framework NOP C++ 4.0. Although the results of the experiment carried out were not promising in terms of performance, with the implementation in the MBRIS 0.1 Framework presenting superior execution times to the implementation of the operations in the PON C++ 4.0 Framework in all scenarios, they demonstrate a technique for implementing the logical-causal flow of NOP notifications through SpMV scenarios.