Towards a Life-Long Learning Soccer Agent

One problem in robotic soccer (and in robotics in general) is to adapt skills and the overall behavior to a changing environment and to hardware improvements. We applied hierarchical reinforcement learning in an SMDP framework learning on all levels simultaneously. As our experiments show, learning simultaneously on the skill level and on the skill selection level is advantageous since it allows for a smooth adaption to a changing environment. Furthermore, the skills we trained turn also out to be quite competitive when run on the real robotic players of the players of our CS Freiburg team.