Multi-approaches to achieve an advanced cognitive agent in a new type of parallel processing computer

Abstract

In this work, we addressed the problem of developing an agent-based artificial general intelligence that can be implemented in compact and power-efficient electronic hardware. We proposed an approach intended to show the feasibility of using this conceptual hardware-based architecture to replicate simple cognitive behaviours. This research started with surveys on cognitive behaviours and their decision-making architectures and compared them with a production rule-based parallel processing computation architecture. In order to demonstrate their potential, a sample case study of the homing behaviour of honey bees was undertaken to demonstrate the possibility of reproducing cognitive behaviours using a production rule-based cognitive architecture. We developed rule-based agents for a mobile platform which, under experimental conditions, made decisions to retrace its path back to a target position by comparison with the reference images. The agent made consistent overall cognitive decisions using fuzzified elements and guided the system reliably to target positions. Then, the research shifted to finding cognitive data representations and constructing cognitive decision-making structures in that production rule-based system. We introduced a new symbolic way of describing the significant features in an image, which is to use a collection of fuzzy symbolic elements to describe the characteristics of the current environmental information. It filtered out any unnecessary details, yet retained sufficient information describing the frame to enable reliable comparisons between images for the purposes of navigation. Numerical data were converted into fuzzy symbolic representations of the surrounding environment. The modified Fuzzy Inference System includes the reasoning rules used to support the cognitive decision-making process. One of the main disadvantages of a rule-based approach is the effort spent on developing rules. In order to reduce the workload of developing rules manually for agents, a modified Association Rules Mining (ARM) method was introduced to discover effective rules for agents autonomously, based on training data sets. This novel rule development method has been demonstrated through a trainable autonomous parking system, which can develop rules for autonomous parking agents.