[topological_prediction] prediction is skewed by many 'fatal' statistics

[bfalacerda]

So, I was looking at some TSC data from last week. The robot tried an edge 4 times. The three first times is succeeded, all on Monday morning, between 10 and 12h

Then Wednesday afternoon, if failed. Since topological navigation retries, and all attempts are added to the nav_stats collection, we have the following 103 failures:

Wednesday, January 18 2017, at 17:18:09 hours
Wednesday, January 18 2017, at 17:18:11 hours
Wednesday, January 18 2017, at 17:18:12 hours
Wednesday, January 18 2017, at 17:18:13 hours
Wednesday, January 18 2017, at 17:18:14 hours
Wednesday, January 18 2017, at 17:18:15 hours
Wednesday, January 18 2017, at 17:18:16 hours
Wednesday, January 18 2017, at 17:18:18 hours
Wednesday, January 18 2017, at 17:18:24 hours
Wednesday, January 18 2017, at 17:18:25 hours
Wednesday, January 18 2017, at 17:18:27 hours
Wednesday, January 18 2017, at 17:18:28 hours
Wednesday, January 18 2017, at 17:18:29 hours
Wednesday, January 18 2017, at 17:18:30 hours
Wednesday, January 18 2017, at 17:18:35 hours
Wednesday, January 18 2017, at 17:18:36 hours
Wednesday, January 18 2017, at 17:18:37 hours
Wednesday, January 18 2017, at 17:18:39 hours
Wednesday, January 18 2017, at 17:18:40 hours
Wednesday, January 18 2017, at 17:18:41 hours
Wednesday, January 18 2017, at 17:18:44 hours
Wednesday, January 18 2017, at 17:21:09 hours
Wednesday, January 18 2017, at 17:21:11 hours
Wednesday, January 18 2017, at 17:21:12 hours
Wednesday, January 18 2017, at 17:21:13 hours
Wednesday, January 18 2017, at 17:21:14 hours
Wednesday, January 18 2017, at 17:21:16 hours
Wednesday, January 18 2017, at 17:21:19 hours
Wednesday, January 18 2017, at 17:27:17 hours
Wednesday, January 18 2017, at 17:27:19 hours
Wednesday, January 18 2017, at 17:27:21 hours
Wednesday, January 18 2017, at 17:27:22 hours
Wednesday, January 18 2017, at 17:27:26 hours
Wednesday, January 18 2017, at 17:27:30 hours
Wednesday, January 18 2017, at 17:27:31 hours
Wednesday, January 18 2017, at 17:27:32 hours
Wednesday, January 18 2017, at 17:27:33 hours
Wednesday, January 18 2017, at 17:27:38 hours
Wednesday, January 18 2017, at 17:27:39 hours
Wednesday, January 18 2017, at 17:27:40 hours
Wednesday, January 18 2017, at 17:27:41 hours
Wednesday, January 18 2017, at 17:27:43 hours
Wednesday, January 18 2017, at 17:27:49 hours
Wednesday, January 18 2017, at 17:27:51 hours
Wednesday, January 18 2017, at 17:27:52 hours
Wednesday, January 18 2017, at 17:27:53 hours
Wednesday, January 18 2017, at 17:27:59 hours
Wednesday, January 18 2017, at 17:28:00 hours
Wednesday, January 18 2017, at 17:28:02 hours
Wednesday, January 18 2017, at 17:28:03 hours
Wednesday, January 18 2017, at 17:28:09 hours
Wednesday, January 18 2017, at 17:28:10 hours
Wednesday, January 18 2017, at 17:28:11 hours
Wednesday, January 18 2017, at 17:28:12 hours
Wednesday, January 18 2017, at 17:28:19 hours
Wednesday, January 18 2017, at 17:28:20 hours
Wednesday, January 18 2017, at 17:28:22 hours
Wednesday, January 18 2017, at 17:28:23 hours
Wednesday, January 18 2017, at 17:28:30 hours
Wednesday, January 18 2017, at 17:28:31 hours
Wednesday, January 18 2017, at 17:28:32 hours
Wednesday, January 18 2017, at 17:28:34 hours
Wednesday, January 18 2017, at 17:28:39 hours
Wednesday, January 18 2017, at 17:28:40 hours
Wednesday, January 18 2017, at 17:28:41 hours
Wednesday, January 18 2017, at 17:28:43 hours
Wednesday, January 18 2017, at 17:28:50 hours
Wednesday, January 18 2017, at 17:28:51 hours
Wednesday, January 18 2017, at 17:28:52 hours
Wednesday, January 18 2017, at 17:28:54 hours
Wednesday, January 18 2017, at 17:28:59 hours
Wednesday, January 18 2017, at 17:29:00 hours
Wednesday, January 18 2017, at 17:29:02 hours
Wednesday, January 18 2017, at 17:29:03 hours
Wednesday, January 18 2017, at 17:29:10 hours
Wednesday, January 18 2017, at 17:29:11 hours
Wednesday, January 18 2017, at 17:29:12 hours
Wednesday, January 18 2017, at 17:29:13 hours
Wednesday, January 18 2017, at 17:29:19 hours
Wednesday, January 18 2017, at 17:29:20 hours
Wednesday, January 18 2017, at 17:29:22 hours
Wednesday, January 18 2017, at 17:29:23 hours
Wednesday, January 18 2017, at 17:29:29 hours
Wednesday, January 18 2017, at 17:29:31 hours
Wednesday, January 18 2017, at 17:29:32 hours
Wednesday, January 18 2017, at 17:29:33 hours
Wednesday, January 18 2017, at 17:29:39 hours
Wednesday, January 18 2017, at 17:29:40 hours
Wednesday, January 18 2017, at 17:29:42 hours
Wednesday, January 18 2017, at 17:29:43 hours
Wednesday, January 18 2017, at 17:29:50 hours
Wednesday, January 18 2017, at 17:29:51 hours
Wednesday, January 18 2017, at 17:29:53 hours
Wednesday, January 18 2017, at 17:29:54 hours
Wednesday, January 18 2017, at 17:30:00 hours
Wednesday, January 18 2017, at 17:30:01 hours
Wednesday, January 18 2017, at 17:30:03 hours
Wednesday, January 18 2017, at 17:30:04 hours
Wednesday, January 18 2017, at 17:30:10 hours
Wednesday, January 18 2017, at 17:30:11 hours
Wednesday, January 18 2017, at 17:30:13 hours
Wednesday, January 18 2017, at 17:30:14 hours
Wednesday, January 18 2017, at 17:32:55 hours

Of course then the prediction for successfully going through this edge is less than 0.05. We should change the statistics so that only one entry is added in these situations

[gestom]

The problem here is that all these events are considered to be independent of each other. To deal with that, one would need to modify the FreMEn (or any probabilistic model) to consider the individual events being not independent. I am not so sure how to do that. @bfalacerda how do you solve that in your case? I mean, how do you deal with the fact that the probablilities of edges' traversals along a planned path are not independent and the probability of traversing a path is different than a simple product of the probabilities of traversing the paths' edges ?

[bfalacerda]

I dont have any elegant way to deal with it. My suggestion is to identify these dependent events during execution (checking for consecutive fails in the same edge within some time threshold I guess) and not add those stats

[Jailander]

Hi, yes what we were thinking is that tampering with the stats is not the best approach, so we are trying to address this on the prediction, we have an idea but needs testing, check gitter for more details

[gestom]

Bruno, its the edge that ends at the charging station, right? In such case, the predictions indicate that the probability of charging is close to 0, which, if I correctly interpret the G4S chat, is a correct prediction.