We propose a sequential learning algorithm with a focus on robot
control. It is initialised by a teacher who directs the robot
through a series of example solutions of a problem. Left alone, the
control chooses its next action by prediction based on a variable
order Markov chain model selected to minimise a MDL criterion based
on generalised code length L_a of the past robot-environment
interaction. The user specifies the parameter a and as a
result, the robot can be directed towards exploratory behaviour if
confidence in the teacher is low (a<0), and towards
goal-seeking exploitive behaviour if confidence in the teacher is
high (a>0). The novelty of the proposed method lies in the
use of generalised code length in the MDL model selection criterion.