Conclusion and Future Work

This paper has presented the layered learning paradigm and illustrated it with a fully-implemented example in the robotic soccer domain. Our layered learning implementation, along with robust low-level skills and a sophisticated team member agent architectures which incorporates a flexible teamwork structure [Stone & Veloso, 1999a], has contributed to the success of our complete team of simulated robotic soccer competitions.

An important direction for future work is to apply layered learning in a new domain. As an example apparently orthogonal to robotic soccer, consider natural language understanding as another application of layered learning. Natural language understanding can have a clear hierarchical task decomposition. For example, learned word sense disambiguation could facilitate learned sentence parsing, which in turn could facilitate semantic encoding of sentences or paragraphs (see Table 5). While it is currently not possible

 

Table: Natural language understanding: a second proposed layered learning application.

Layer	Learning Task
L1	Word sense disambiguation
L2	Sentence syntax
L3	Sentence semantics

 

in general to learn sentence semantics straight from a string of words, a hierarchical decomposition of the task coupled with the layered learning paradigm may render the learning task tractable.

Layered learning is potentially applicable to any complex learning problem for which a hierarchical decomposition exists. Its power is derived from the concept of directly combining different ML algorithms within a hierarchically decomposed task representation.