Developmental Learning

Stream 2b: Developmental Learning

Flexible real-world problem solving often requires sensitivity to subtle task and object related features. In developing artificially intelligent “thinking systems”, it is doubtful that such subsymbolic sensitivities can be hand-coded or learned through explicit tuition. Instead, learning appropriate grounded representations through interactions with and exploration of the world is an important (and perhaps necessary) characteristic of artificially intelligent embodied systems. Stream 2b of this project will continue the work of Bolland and Emami (2007), in exploring possible neurologically inspired mechanisms that may lead to the emergence of competence in embodied systems, and the self-acquisition of subtle real-world knowledge that is crucial for high-level problem solving.

The adjacent video demonstrates our initial foray into the area of developmental robotics. The Aibo has a set of basic reflexes such as hitting, turning and biting, but no innate knowledge of what effect this has on the world.

With initial learning, the AIBO develops predictions about what is likely to occur – which, initially, is not much. However, when this expectation is violated (such as when an object makes a noise or moves as a result of the action), this behaviour is reinforced. In humans, such “expectation violations” are central in the self-acquisition of knowledge, releasing dopamine in the brain - a neurotransmitter that shapes behaviour and is associated with a feeling of “pleasure.”

Home - Stream 1 - Stream 2a - Stream 2b - Stream 2c