Topics to work on

[Erfan7bt]

• heatmap of the clusters
• better visualizing of path plots
• difficulty score
• switching strategies analysis

[Erfan7bt]

heatmap of the clusters implemented for all pnp puzzles in: https://github.com/jinanallan/puzzlesdata/commit/568311b54262e1fbdd7716e79b9429d59862dcd1

a better visulizaztion of path plots based on frame file tested here: https://github.com/jinanallan/puzzlesdata/commit/4b24f8d61781911a2255af0e0411ddbea0935567

an initial implement of diff score also has been tested: https://github.com/jinanallan/puzzlesdata/commit/eeac1bf000aefd6bb4764556f5f296cc0e8abe4a

[Erfan7bt]

The effect of choice of number of clusters need to be discussed more

[Erfan7bt]

after plotting previous results without ego trajectories: one thing seemed interesting, When attached to an objects the trajectories are much smoother compared to ego free movement trajectories, which can have a behavioral reason behind it .

[Erfan7bt]

https://tubcloud.tu-berlin.de/apps/files/?dir=/Shared/scioi_slides&openfile=2034382098

Marc's slide for the project

[Erfan7bt]

• test the clustering by only dtw calculation for ego
• merge similar puzzle solutions and do the clustering

[Erfan7bt]

Marc's meeting:

• Wrap the time for clusters and generate avg. time series: https://www.sciencedirect.com/science/article/pii/S003132031000453X?via%3Dihub
• https://www.sciencedirect.com/science/article/pii/S030439751100822X?via%3Dihub and making the solutions the same time length
• compute dtw score after ignoring unattached ego movements
• linear combination of states

[Erfan7bt]

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5033575/

[Erfan7bt]

https://ieeexplore.ieee.org/abstract/document/1163055 DTW paper

[Erfan7bt]

https://arxiv.org/pdf/1703.01541.pdf soft DTW barycenter

[Erfan7bt]

Implement soft dtw score with torch and then redo the clustering task https://tslearn.readthedocs.io/en/stable/gen_modules/metrics/tslearn.metrics.soft_dtw.html https://tslearn.readthedocs.io/en/stable/gen_modules/clustering/tslearn.clustering.TimeSeriesKMeans.html#tslearn.clustering.TimeSeriesKMeans

[Erfan7bt]

https://proceedings.mlr.press/v130/blondel21a/blondel21a.pdf soft-DTW divergence is introduced to counteract the non-positivity of soft-DTW:

[Erfan7bt]

https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=994784 good paper on use of stw or LCS and hierarchical clustering

[Erfan7bt]

some new measurements to do :

• effect of preprocessing on max, Avg, or fraction of Neg. or below Avg silhouette score
• effect of $\gamma$ on max, Avg, or fraction of Neg. or below Avg silhouette score

[svetlanalevit]

Motivation to consider positional data for clustering:

• some solutions have same discrete action skeleton, but vary in trajectores : could not be separated
• some solutions include several interactions with the same object, which correspond to the same action, but are executed in an interrupted way, such interactions could be merged Advantages to using both positional and discrete events:
• would allow to merge many discrete actions with similar trajecties, but still distinguish different grasps and changed movement paths (was moving the object up, then regrasped and started moving left: two "actions")

Why do we evaluate clusters:

• justify that we use them to extract strategies
• to see if clustering is robus enough to produce meaningful clusters on new data

[svetlanalevit]

possible claim/idea in the paper:

• mostly on analysis of beh. data
• considering different representations for manipulation puzzles
• discrete events are good for some, positional data for others: how can we combine them?

splitting the technical part of the analysis as a separate publication goal: putting on arxiv first, asking Maartens

[Erfan7bt]

https://www.cs.toronto.edu/~sheila/2542/s14/A1/introtopddl2.pdf pddl