A biologically inspired cognitive intelligence with stress based modelling and working memory optimisation for robot partners

Abstract

This thesis addresses one of the fundamental questions for any ageing society: Who will care for the elderly? The “who” is shaping up to be the highly advanced technology in the area of humanoid robotics. The advantage of humanoid robots is that they can be designed to look and act like humans, but how do we avoid “uncanny valley”, where people reject humanoid robots in the long run for “looking human”, but not “behaving human”? It raises an important question of how we make robots that behave more humanlike, so our ageing societies readily accept them. Critically, how do we advance a robot’s embodied cognitive intelligence? We identify three key components behind the idea of embodied cognitive intelligence: Spiking Neural Network (SNN), working memory and stress response system. The stress response system acts to moderate working memory function, and together they help to encode and summarise a robot’s current environment context information which can be used to optimise available actions, plans and intentions. Meanwhile, the SNN activation will determine the timing of new intuition creation. The ability to understand something immediately, without the need for conscious reasoning. In our model, the outputs of the SNN are defined as intuition. In the first part of this thesis, we explore the idea of working memory and how it is regulated and optimised within biological systems in a Markov Decision Process (MDP) navigation problem. One of the key theories in this area is from Lupien, who demonstrated that stress hormones play a critical role in governing the function of working memory during some uncertainty in the environment. We develop a working memory model based on their stress theories. Our model allows the robot to “focus” on a particular set of actions at any given moment according to the currently perceived environment context information. It enables our model to address the combinatorial explosion problem normally associated with large action sets and allows us to extend standard Q-Learning techniques to enable the robot to select near-optimal actions in real time. In the second part, we address another working memory characteristic, first identified by the cognitive psychologist Braver that he termed dual mechanisms of cognitive control. Proactive control is one of the cognitive control mechanisms whereby a robot’s environment context information is fed back into working memory as an input to enable active working memory optimisation towards some intention. We model this proactive control behaviour using two novels biologically inspired genetic algorithm approaches. Next, we validate our proposed spiking reflective processing model that leads to agent’s creation of new intuition with established psychology tests. We validate our model with an interactive human-robot conversation scenario. Our survey from the human-robot interactive scenarios shows that the proposed model is evaluated to be more human-like in behaviour during human-robot interaction and a few steps closer to our ultimate goal.