| BLEARN | Learning Relational Concepts from Sensor Data of Mobile Robots | Modelling, Diagnosis, Control | b D, Y |
| Klingspor, C., Morik, K. and Rieger, A. Learning concepts from sensor data of a mobile robot. Machine Learning, 1996, to appear
Morik, K. and Rieger, A. Learning action-oriented perceptual features for robot navigation. In Proc. of the 1st European Workshop on Learning Robots, 1993
|
| This set of datasets provides the results of experiments with a real mobile robot described in detail e.g. in [Klingspor 96, Morik 93]. Data have been collected with the goal to learn abstract operational concepts, e.g. robot "moves along a wall", from sequences of sonar sensor measurements and robot actions. Each dataset is represented in the first order logic, when the following restrictions are applied: facts can be linked using the argument of type TIME, and there are never two different facts concerning the same sensor and the same point in time. Positive and negative examples are characterized e.g. through the sensor feature predicates structured according to a general pattern {sf}(Tr,S_id,Start,End,Rel_or). Its intended interpretation is "In a trace Tr, a specific sensor S_id percieved an object corresponding to the predicate's name {sf}. This happened during the time interval between Start and End while moving in a relative orientation Rel_or along the object."
Each data set corresponds to learning disjoint concepts at one level. The levels are organized in a hierarchy as shown below:
high-level concepts
| perception-integrating actions
| perceptual features | | sensorgroup features | / | / sensor features | / | / basic perceptual features | sclass, | basic-actions, | dXsucc raw sensor data period-of-time-perceptions pdirections
Each node in the hierarchy denotes a set of predicates. The links are directed from bottom to top. They link the sets of predicates in nodes of lower level to a set of predicates in a node of higher level, if the predicates of the lower level are necessary to learn the concepts of the higher level. Hence, a sequence of learning passes can learn high-level concepts from raw sensor data.
|