Artificial Agent Language Development based on the Xzistor Mathematical
Model of Mind
Abstract
The Xzistor Mathematical Model of Mind is a cognitive architecture that
uses a functional model of emotions, based on biological homeostatic and
allostatic control loops, to explain how an artificial agent can
systematically learn to navigate to a reward source. The model also
explains how subtasks required for the agent to access reward sources
can be learnt through reinforcement learning. Simple virtual and
physical agent implementations of the model have demonstrated how agents
successfully learn to navigate to reward sources from anywhere in their
environments. These implementations also showed how agents become
motivated to perform subtasks to gain access to the reward sources. This
paper describes how this demonstrated learning ability in agents,
provided by the Xzistor brain model, could be used as a theoretical
basis for implementing a human-like language learning skill in agents.
This goes beyond Large Language Model approaches by incorporating
computational equivalents of many human brain functions, including
sensing, recognition, inference, and emotions. The study concludes that
this cognitive architecture could provide a proof-of-concept
implementation in agents of the principles of verbal behavior identified
by B.F. Skinner (Skinner, 1957). A multi-stage project is proposed to
demonstrate how an artificial Xzistor agent could systematically develop
basic language skills using artificial emotions and reinforcement
learning and, over time, refine this skill towards improved syntax and
grammar use. The paper includes sections on the mathematical principles
underpinning the Xzistor brain model and how it could potentially unify
behaviorist and structuralist language theories.