The behavior of reinforcement learning (RL) algorithms is best underst
ood in completely observable, discrete-time controlled Markov chains w
ith finite state and action spaces. In contrast, robot-learning domain
s are inherently continuous both in time and space, and moreover are p
artially observable. Here we suggest a systematic approach to solve su
ch problems in which the available qualitative and quantitative knowle
dge is used to reduce the complexity of learning task. The steps of th
e design process are to: (i) decompose the task into subtasks using th
e qualitative knowledge at hand; (ii) design local controllers to solv
e the subtasks using the available quantitative knowledge, and (iii) l
earn a coordination of these controllers by means of reinforcement lea
rning. It is argued that the approach enables fast, semi-automatic, bu
t still high quality robot-control as no fine-tuning of the local cont
rollers is needed. The approach was verified on a non-trivial real-lif
e robot task. Several RL algorithms were compared by ANOVA and it was
found that the model-based approach worked significantly better than t
he model-free approach. The learnt switching strategy performed compar
ably to a handcrafted version. Moreover, the learnt strategy seemed to
exploit certain properties of the environment which were not foreseen
in advance, thus supporting the view that adaptive algorithms are adv
antageous to nonadaptive ones in complex environments.