Two timescale immunological learning for idiotypic behaviour mediation

Short-term learning can be defined as taking place over the lifetime of an individual, whereas long-term learning can be the net evolutionary result of knowledge accumulation. The integration of these approaches provides an advantage over the use of either one alone. Our immune system combines both effortlessly through the use of gene libraries and other evolutionary and immunological mechanisms. We aim to imitate this to derive useful rules to autonomously solve navigation / control problems. Without long-term learning, the controller initially lacks knowledge of the effectiveness of rules and therefore has no sense of which ones to use. The disadvantage of long-term learning in isolation is that it must be conducted offline in simulation, due to the time scales involved. The use of simulators can mean that results translate poorly to the physical world. Hence we provide feedback from real short-term experiments to validate the data. We propose here to investigate the relationship between short- and long -term learning by designing a robot controller that considers both learning cycles and allows their outputs to feed into each other. The chosen architecture is the creation of initial rules through an artificial gene library and their accelerated (simulated) evolution to establish a rule database, indexed by actuator function. A subset of the derived rules is transferred to a physical robot that continually adapts and re-selects them using the dynamics of an Artificial Immune System, with rules interconnected through an idiotypic network. The methodology thus provides a set of starting rules, a measure of their usefulness, a means for updating and thus improving the repertoire and a mechanism for adaptive rule selection.