TINLAP '75 Proceedings of the 1975 workshop on Theoretical issues in natural language processing
| Hi-index | 0.00 |
A model of robot perception learning is described that duplicates aspects of observed animal performance. The designer incorporates in the model abstract knowledge about the possible observable consequences of actions in the robot's world. In this way the robot can predict an outcome and can compare the expectation with the actual experience. The robot's internalized perceptions of the world may then be linked with the consequences. The model structure permits the robot to reinforce predictions if they are in accord with experience or augment its stored perception knowledge and predictive ability if the predicted consequences disagree with expectation. The learning is accomplished without extensive search over a multi-dimensional space. The model is being implemented as a system called RECOGNIZER, intended to work with situations actually arising in the world of the JPL robot, now operational.