A Method for Clustering the Experiences of a Mobile Robot that Accords with Human Judgments Tim Oates, Matthew D. Schmill, Paul R. Cohen Discussant: Jacek Rawicki
Introduction A robot agent Unsupervised method of learning Individual time series Clustering: -Measure of similarity -Cluster prototypes
Clustering Experiences Obtaining data (experiences) Dynamic Time Warping (DTW) Distance matrix Cluster prototypes
Evaluation High levels of accordance, but: Ordering effect (greedy clustering algorithm) Errors from DTW (manipulation of time dimension)
Conclusion High level of accordance A simple optimization technique My ( humble : – ) suggestions: Use DDTW instead of DTW Use other clustering techniques
Errors in DTW a pathological result: the algorithm tries to explain variability in the Y-axis by warping the X-axis.
Improvement of DTW - DDTW Derivative Dynamic Time Warping takes into consideration the first derivative.
DTW (above) vs. DDTW (below)
References DDTW: Keogh E. J., Pazzani M. J.: Derivative Dynamic Time Warping, Other clustering techniques: e.g. Jain A.K., Murty M. N., Flynn P. J. Data Clustering: A Review