De Jager, Deon (2020) Real-time optimization of working memory in autonomous reasoning for high-level control of cognitive robots deployed in dynamic environments. (PhD thesis), Kingston University, .
Abstract
High-level, real-time mission control of autonomous and semi-autonomous robots, deployed in remote and dynamic environments, remains a research challenge. Robots operating in these environments require some cognitive ability, provided by a simple, but robust, cognitive architecture. The most important process in a cognitive architecture is the working memory, with core functions being memory representation, memory recall, action selection and action execution, performed by the central executive. The cognitive reasoning process uses a memory representation, based on state flows, governed by state transitions with simple, quantified propositional transition formulae. In this thesis, real-time working memory quantification and optimization is performed using a novel adaptive entropy-based fitness quantification (AEFQ) algorithm and particle swarm optimization (PSO), respectively. A cognitive architecture, using an improved set-based PSO is developed for real-time, high-level control of single-task robots and a novel coalitional games-theoretic PSO (CG-PSO) algorithm extends the cognitive architecture for real-time, high-level control in multi-task robots. The performance of the cognitive architecture is evaluated by simulation, where a UAV executesfour use cases: Firstly, for real-time high-level, single-task control: 1) relocating the UAV to a charging station and 2) collecting and delivering medical equipment. Performance is measured by inspecting the success and completeness of the mission and the accuracy of autonomous flight control. Secondly, for real-time high-level control of multi-task autonomous vehicle control: 3) delivering medical equipment to an incident and 4) provide aerial security surveillance support. The performance of the architecture is measured in terms of completeness and cognitive processing time and cue processing time. The results show that coalitions correctly represent optimal memory and action selection in real-time, while overall processing time is within a feasible time limit, arbitrarily set to 2 seconds in this study.
Actions (Repository Editors)
 |
Item Control Page |