Hybrid System Model

Some systems based on translational architectures can also provide both translation methods: from rules to network and from networks to rules. These systems are usually applied to perform theory refinement [Towell91, Fu93, Mahoney96, Osorio99].

Systems with both translation methods can be easily integrated with a symbolic inference engine. Those architectures composed of more than one module, incorporating complete symbolic and complete connectionist components (symbolic and connectionist modules cooperation), are classified under hybrid architecture approach. They are also named hybrid modular architectures and functional hybrids. Translational architectures are a degenerate case of hybrid modular architectures.

Hybrid modular architectures can be classified according to the degree of integration and the mode of integration [Melanie97, Cloete00, Wermter00]. The degree of integration is a criterion used to measure the level of interaction between the two modules. This can be loosely coupled, tightly coupled, or fully integrated.

The mode of integration may be classified as chain processing, sub-processing, meta-processing, or co-processing. In the chain processing mode, the symbolic module acts as the main problem solver and is assisted by a neural pre-processor and/or a post-processor, or, the neural module is the main processor and is assisted by a symbolic pre-processor and/or post-processor. In sub-processing, one of the two components is embedded in and subordinated to the other, which acts as the main problem solver. In meta-processing, one module acts to solve the problem and the other one act on the “meta-level” (e.g., supervision, performance control, error detection) in relation to the first one. In the last mode, co-processing, both components are equal partners in the problem-solving process: each can interact directly with environment; each can transmit information to and receive information from the other.

The second generation ES’s, supported by hybrid knowledge-based neurocomputing, exploit the combination of symbolic and connectionist modules allowing us to improve the final system performance, to handle uncertain and inexact knowledge, and also to integrate knowledge obtained from different sources (e.g., experts and databases).