Connectionist expert systems construct their knowledge bases and conclude with the help of artificial neural networks.

In connectionist expert systems, inputs take the form of text or other naturally occurring data, and an artificial neural network (ANN) then creates inferencing rules to apply. In addition, rough set theory can be utilised to improve the knowledge encoding in the weights, and evolutionary algorithms can be employed to enhance the search optimisation. 

Evolution of connectionist models

The popularity of connectionist models in AI has swung from tremendous enthusiasm in the 1960s to a complete taboo in the 1970s. Nevertheless, there is a surge of interest in these approaches: we estimate that research in this area has expanded by two orders of magnitude in the last five years. Theoretical breakthroughs caused these swings in interest. Many researchers were intrigued by the early success of neuron-like models termed perceptrons in learning. Then, in 1969, Minsky and Papert identified significant limits of perceptrons, leading to the widespread abandonment of this line of research. 

Recent improvements have helped to avoid or overcome many of the flaws in learning, resulting in the current flourishing of connectionist activity. Psychologists are interested in connectionist models because of their structural and behavioural similarities to neuronal systems. In contrast, AI researchers desire the ability to design algorithms for machine learning, which is one of the fundamental concerns in AI.

Human behaviour replication

Connectionist network techniques have been tried to replicate human behaviour in visual and vocal learning and recognition domains with variable degrees of success. In addition, human reasoning has been simulated using logic (or "symbolic") techniques. However, there is a significant schism between AI and cognitive psychology; both approaches rarely address the same topic. 

Suppose a connectionist network can do the same task as a rule-based system. Studying the network's internal representations may provide insights into the "microstructure" of how reasoning processes occur in the human brain. These are the same reasoning processes that expert system rules outline at a higher description level. However, people are still determining how this type of effort will affect the divide between the symbolic and non-symbolic factions (stated in the introduction). 

Biological validity

While connectionist networks have considerable biological validity, it is hard to conceptualise a neural basis for the back-propagation method. But why should engineers worry about that? As a group, they care less about understanding how people think than they do about putting their ideas into action. As an example, consider the recent market boom in expert system software. Despite the allure, implementing an expert system is typically time-consuming and costly.

As practitioners define it, forecasting is a unique combination of informal reasoning within very soft restrictions offered by frequently incomplete and erroneous data. Because of the type of reasoning needed, classic rule-based techniques are a logical fit. However, data on solar and flare occurrences are often unreliable and noisy. As a result of the nature of the data, careful handling of rule strengths and certainty factors is required. However, dealing with this data type is one of the purported advantages of connectionist networks. Furthermore, some of the reasoning involves pattern matching across multiple data types. It prompted us to believe that a connectionist network could learn the essential internal representations to handle this task.

Image source: Unsplash