Adaptation of the Perception-Action Loop Using Active Channel Sampling

Abstract

During the lifetime of a real world agent or robot, many changes unforeseen at design time can occur. Whether these are due to a change in environmental conditions or to alterations of the embodiment of the robot, flexibility and adaptation are essential qualities that can help it to keep operating in this new situation. This work is based on an information-theoretic approach and introduces an exploration strategy that allows an agent to detect and adapt to changes in its perception-action loop by actively sampling areas of interest. We define the problem of exploring the sensorimotor channel and establish a measure of the distance between the observed and the real model of the channel. An optimal Oracle-based strategy is used to compare performances of the adaptive sampling strategy and a random strategy. Results for different scenarios of change in a binary channel show that the proposed strategy is highly effective in many cases. We also outline principles to adapt this mechanism to the exploration of multiple channels and we give preliminary results for such a scenario.

A distributed stochastic perception-action loop model of cell motility

“…In the context of robotics and artificial life, perception-action loops are generally viewed as an arbitrarily complex mapping from sensory input to a motor output [1], [2]. The sensory input is a representation of the environment, which, through filters and classifiers of any possible form, is processed into a motor action.…”

Effects of Anticipation in Individually Motivated Behaviour on Survival and Control in a Multi-Agent Scenario with Resource Constraints

“…Capdepuy et al [28] proposed an information-theoretic mechanism to create an internal representation of the agent's environment, and subsequently, they developed an information-theoretic anticipation framework to identify relevant relationships between events [29]. They also evaluated an active exploration strategy, in which an agent gradually collects samples of the interaction with the environment, such that the prediction accuracy becomes maximized [30].…”