1. Title: Learning Concepts from Sensor Data of a Mobile Robot 2. Source Information (a) Donors: Volker Klingspor, Katharina J. Morik, Anke D. Rieger Computer Science Dept. LS VIII University of Dortmund, Germany (b) Date: 15.07.95 3. Past Usage Learning Concepts from Sensor Data of a Mobile Robot, to appear in: Machine Learning Journal, 1995 4. Relevant Information We provide here a set of data sets, where 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, | succ raw sensor data period-of-time-perceptions pdirections Hence, a sequence of learning passes can learn high-level concepts from raw sensor data. disjoint goal predicates: background predicates: 0. (basic perceptual features): measurement(Tr,T,S,Dist,Sx,Sy,Sa,Obj,E) stable(Tr,Or,S,T1,T2,Gr), increasing/6, decreasing/6, incr_peak/6, decr_peak/6, no_movement/6, no_measurement/6, something_happened/6, straight_away/6, straight_to/6 1. (sensor features): basic perceptual features s_line(Tr,S,T1,T2,RM), s_jump/5, s_convex/5, s_concave/5 2. (sensorgroup features): sensor features, sg_line(Tr,S_C,T1,T2,Mv), sclass(Tr,S,T1,T2,S_C), sg_jump/5, sg_convex/5, sg_concave/5 succ(T1,T2) 3. (perceptual features): sensorgroup features, through_door(Tr,T1,T2,Mv) succ(T1,T2) 4. (perception-integrating actions) basic actions, standing(Tr,T1,T2,Speed,MD,Perc,PD), perceptual features, moving/7, parallel-moving/7, period-of-time-perceptions rotating/7, 5. (high level concepts) perception-integrating actions, move_through_door(Tr,T1,T2,T3), pdirections rotate_in_front_of_door/4, move_in_front_of_door/4, move_along_door/4, move_parallel_in_corner/4, rotate_in_front_of_wall/4, move_closer_to_wall/4, rotate_in_corner_to_front_wall/4 The files needed to perform the learning passes are listed in this tabular. Note, that the acquisition of basic perceptual features from raw sensor data (predicate: measurement) requires some calculations. The data set for learning basic features is NOT of the regular form, but presents the sensordata without classification. Each data set (except for the basic features) is of the form: 1. instances of target predicate 2. background knowledge about classified instances 3. general background knowledge. _____________________________________________________________________________________________ |Files for: goal predicate: |instances background additional background ________________________________|____________________________________________________________ | basic perceptual features: | all basic perceptual features | Measurements | sensor features: | s_concave |S_concave.ex BF_set1 or BF_set2 or BF_set3 s_convex |S_convex.ex BF_set1 or BF_set2 or BF_set3 selected traces: | (t30|t34|t35|t100|t101|t5|t6|t7)| (different data sets for the same goal predicate were (t30|t34|t35|t100|t101|t5|t6|t8)| taken from the specified files by using a unix command (t30|t34|t35|t100|t101|t5|t7|t9)| egrep -e "(t30|...)" filename (eg. S-convex.ex, BF_set1) | s_jump |S_jump.ex BF_set1 or BF_set2 or BF_set3 selected traces ('egrep'): | (t30|t34|t35|t100|t101|t5|t6) | (t30|t34|t35|t100|t101|t5|t8) | (t30|t34|t35|t100|t101|t7|t9) | | s_line |S_line.ex BF_set1 or BF_set2 or BF_set3 selected traces ('egrep'): | (t30|t34|t35|t100|t101|t5) | (t30|t34|t35|t100|t101|t7) | (t30|t34|t35|t100|t101|t9) | | sensorgroup features: | (no trace selection) sg_convex |SG_convex.ex S_convex.ex sclass_123.bg, succ.bg sg_concave |SG_concave.ex S_concave.ex sclass_123.bg, succ.bg sg_line |SG_line.ex S_line.ex sclass_123.bg, succ.bg sg_jump |SG_jump.ex S_jump.ex sclass_123.bg, succ.bg selected traces ('egrep'): | (t30|t34|t35|t100|t101|t5|t6|t7)| (t30|t34|t35|t100|t101|t5|t6|t8)| (t30|t34|t35|t100|t101|t5|t7|t9)| | | perceptual features: | through_door |Door.ex SG.ex | | perception integrating actions:| all actions |Piaf-all.ex Percfeat, BasicActs time-perc.bg | operational concepts: | all operational concepts |Opc-all.ex Piaf-all.ex pdirections.bg _____________________________________________________________________________________________ the above files are packed together in tar-files (use tar xvf [name] to extract) Bfsets.tar: Measurements, BF_set1,2,3 Sfeatures.tar: S_concave.ex, S_convex.ex, S_line.ex, S_jump.ex SGfeatures.tar: SG_concave.ex, SG_convex.ex, SG_line.ex, SG_jump.ex, sclass_123.bg, succ.bg Pfeatures.tar: Door.ex, SG.ex Higherlevels.tar: Opc-all.ex, Piaf-all.ex, pdirections.bg, Percfeat, BasicActs, time-perc.bg 6. Sort information sorts: Tr (Trace) (integer) T (Time) (integer) S (Sensor) (integer 0-23) Or (Orientation) (real 0-360) Sa (S-Orientation) (real 0-360) Gr (Gradient) (real) Dist (Distance) (real) Sx,Sy (Sensor-coordinates) (real) Obj (Object) (integer) E (Edge) (integer) S_C (Sensorclass) (set of front_side,right_side,back_side,left_side ...) Mv (Movement) (set of parallel, diagonal) MD (MoveDirection) (set of forward, backward, right, left) PD (PerceptionDir.) (set of front, rear, right, left) Perc (perceptual features)