Emergent Structuring of Interdependent Affordance Learning Tasks Using Intrinsic Motivation and Empirical Feature Selection

This paper studies mechanisms that produce hierarchical structuring of affordance learning tasks of different levels of complexity. Guided by intrinsic motivation, our system detects easy tasks first, and learns them in selected environments which are maximally different from the previously encountered ones. Easy tasks are learned from observed low-level attributes of the environment, and provide abstractions over these attributes. As learning progresses, the system shifts its focus and starts learning harder tasks not only from low-level attributes but also from previously-learned abstract concepts. Therefore, hard tasks are autonomously placed higher in the hierarchy if the easy task concepts are identified as distinctive input attributes of hard tasks. Use of abstract concepts allows hard tasks to be learned faster than learning them from scratch, i.e., from lowlevel perception only. We tested our system with the tasks of learning effect predictions for poke and stack actions using a dataset that includes 83 real-world objects. On the basis of a large number of runs of the method, our analysis shows that the hierarchical task structure emerged as expected, along with a consistent learning order. Furthermore, a significant bootstrapping effect in learning speed of the stack action was observed with the discovered hierarchy, albeit only when fully-learned poke actions were used from the beginning.