8th July, 2016

Island Heuristics

Current Status

Start-goal pairs : 7-8

Dimensions : 7 angles of the right arm

Though 7-8 pairs are enough for base clustering, I need to use more for 7 dimensions. For max_kitchen, the key heuristic is to fold arms in narrow passages. I first compared island heuristics (IH) with the normal setup (including tuck-in heuristic) and island heuristics perform worse than them. However, IH perform much better than when tuck-in is not being used, implying that meanshift is able to figure out somewhat the right arm states.

Note: I have been using states generated by MHA using the tuck-in heuristic, which means my training data is biased with tucked-in states. However, it is still interesting to note that mean shift is able to extract the useful modes.

What now?

1. Firstly, get more states by running meanshift on more paths (10 is too low). (Done - Leads to better islands)

2. Improve the metric I use to compute the cost to go to an island state. Currently, I am using plain vanilla Euclidean. I can introduce weighted sum of individual differences of angles and take into account the joint angle constraints.

3. Check if gaussian kernel performs better.

4. Use a state-gaph-only version of meanshift.

Results

This is for tableObstacle1 (Island with 5 islands without tuck-in)

total-plan-time initial-solution-planning-time epsilon-1 initial-solution-expansions final-epsilon-planning-time epsilon-2 solution-epsilon expansions solution-cost path-length 186.23 186.23 50.00 1173.00 186.23 50.00 50.00 1173.00 45294.00 579.00

91.48 91.48 50.00 531.00 91.48 50.00 50.00 531.00 26339.00 220.00

127.83 127.83 50.00 942.00 127.83 50.00 50.00 942.00 12927.00 181.00

159.97 159.97 50.00 1147.00 159.97 50.00 50.00 1147.00 36184.00 424.00

300.00 300.00 50.00 2985.00 300.00 50.00 50.00 2985.00 1000000000.00 0.00

85.28 85.28 50.00 447.00 85.28 50.00 50.00 447.00 22348.00 178.00

300.13 300.13 50.00 3687.00 300.13 50.00 50.00 3687.00 1000000000.00 0.00

136.30 136.30 50.00 1263.00 136.30 50.00 50.00 1263.00 14331.00 105.00

300.02 300.02 50.00 3152.00 300.02 50.00 50.00 3152.00 1000000000.00 0.00

12.43 12.43 50.00 62.00 12.43 50.00 50.00 62.00 7131.00 24.00

                   vs

(Default implementation)

122.42 122.42 50.00 2195.00 122.42 50.00 50.00 2195.00 55698.00 573.00

33.61 33.61 50.00 488.00 33.61 50.00 50.00 488.00 13744.00 211.00

37.55 37.55 50.00 640.00 37.55 50.00 50.00 640.00 24216.00 261.00

166.31 166.31 50.00 3095.00 166.31 50.00 50.00 3095.00 57237.00 555.00

46.95 46.95 50.00 728.00 46.95 50.00 50.00 728.00 40808.00 314.00

29.48 29.48 50.00 411.00 29.48 50.00 50.00 411.00 20789.00 189.00

300.01 300.01 50.00 8705.00 300.01 50.00 50.00 8705.00 1000000000.00 0.00

92.99 92.99 50.00 1891.00 92.99 50.00 50.00 1891.00 17710.00 145.00

51.21 51.21 50.00 826.00 51.21 50.00 50.00 826.00 34602.00 480.00

8.53 8.53 50.00 88.00 8.53 50.00 50.00 88.00 3122.00 22.00