Stability of learning in the ARBIB autonomous robot

摘要

We have previously described the ARBIB autonomous robot which consists of a mobile platform running a neural network simulator.Unlike most other behaving robots, the neural system is biologically-inspired and operates at the level of individual spikes.Rather than using currently-popular reinforcement learning techniques, ARBIB learns from exposure to its environment via low-level mechanisms of habituation, sensitisation and classical conditioning.In previous work, its short-term memory (formed through recovery of synaptic weights) made its learning almost totally plastic; it had no long-or medium-term memory.This paper explores means of adding stability to its learning.Long-term memory is provided by a simple form of synaptogenesis, which forms new connections within the nervous system.Medium-term memory is provided by a recurrent neural circuit coupled to a simple model of the medial pallium, which in turn is fed from a sonar range-finding cell.This allows the nervous system to respond to successive stimuli that lie outside the duration of its short-term memory.The effects of these two enhancements are assessed for their impact on ARBIB's behaviour.Long-term memory is tested by collision avoidance behaviour, demonstrated by presenting a comparison of firing activity in bump sensory cells with and without synaptogenesis.Medium-term memory is tested by allowing the sonar-driven medial pallium to habituate to a distant target, and then introducing a transitory target at close range.In this way, a useful measure of learning stability through medium-and long-term memories is achieved.