Human Inspierd Algorithms and Algorithm Discovery Methods for Robotic Intelligence
Vladimirov Blagovest · Institutional Repositories DataBase (IRDB) · 2007
The objective of the presented research is to identify and implement algorithms providing robotic intelligence.The starting point of the research is implementing algorithms for robotic intelligence using a neural network based cognitive control model, since it has the potential to combine the learning capability of neural networks with the flexibility of symbolic systems.To confirm and utilise this potential, we proposed two algorithms inspired by high-level human cognitive abilities.A further step toward practical application requires improvements of the process for obtaining the algorithms' inputs.For that purpose, we proposed two algorithm discovery methods targeting algorithms used in low-level, basic information-processing human abilities.The first proposed algorithm is based on the human ability for mental rehearsal and provides a functionality for fast switching among the learned behaviours according to environment and task requirements.The algorithm is characterised by a neural network implementation of behaviour-based control system that supports learning combinations of basic behaviours, and acquiring internal representations that govern the behaviour switching.Unlike the most approaches that encode the robot's behaviour only into the neural network's weights, here it is partially encoded as a sustained activation of the context layer neurons, which allows faster changes.An expected benefit of implementing the proposed algorithm is providing behaviour flexibility for a control system of a cleaning robot or a pet robot, for example.The second proposed algorithm provides realisation of task specific, human abilities by learning from successful task solutions obtained from skilled individuals.The algorithm is characterised by a modified neural network realisation of a cognitive control model that allows learning from positive examples only.In tests performed using data from human i I would like to thank many people for their support and encouragement during my study.