dlab IRC 2017 October 28th

← Previous day: 2017-10-27 · Next day: 2017-10-29

#dlab logs for 2017-10-28

Current topic
A distributed laboratory. Everything which appears here is logged at https://rawl.es/dlab/irc/2017/
adamfc (Adam Forsythe-Cheasley)
[10:23:29Z] rawles: I have a registered account
rawles (Simon Rawles)
[10:45:57Z] adamfc: I think I've just set things up so you get operator status when you connect to this channel.
[10:48:33Z] If it doesn't work, try: /msg ChanServ op #dlab
[14:37:30Z] supertime: Yeah, as usual I kind of hacked out something that worked first and then realised that it's just topological sort, with a twist. Things are working really nicely now and I can go onto the next bit.
[14:41:23Z] There's probably whole Java frameworks for this, but I kind of want to do it myself so I can extend it easier.
[14:58:39Z] supertime: If you had lots of time, what personal coding/ML/etc project would you do?
supertime (Tim Kovács)
[15:21:59Z] I’d make a rule explorer. When you run XCS it prints out hundreds of rules with many parameters and you stare at them and try to make sense of what it’s done.
[15:22:28Z] I’d make a tool that lets you sort and filter and query and plot them.
[15:22:36Z] How would the rule explorer work, what would be the dialogue between the computer and the user?
[15:22:48Z] Oh, you just answered that.
[15:25:04Z] So what would a query be, 'tell me all the rules that use feature i'?
[15:26:00Z] I’m don’t know what the difference between filters and queries are. Query just sounded like a good word.
[15:26:27Z] But yes to your example. Or ‘show rules with x>5 AND y<10'
[15:27:36Z] That might also be good for ensemble methods in general.
[15:28:25Z] Yeah, some at least. An ensemble of trees for example, since trees are equivalent to rules
[15:29:18Z] Yeah, like if you train a random forest, you know each of the trees are interpretable, but the forest itself can't explain the data as well. It's just too much to take in.
[15:29:37Z] I guess not everyone wants to explain the data.
[15:30:05Z] Yes, the point is to understand the output better.
[15:30:34Z] But besides understanding the data, it helps understand the algorithm and the parameter settings
[15:31:02Z] So you can learn how to tweak it so it gets the kind of rules you want?
[15:31:19Z] Maybe it could automate that process too.
[15:32:06Z] yeah. i kind of learned what to look for in the output, but it’s akward when it’s just in a text file, so the tool would be useful
[15:32:17Z] automating it would be useful
[15:33:40Z] Maybe someone at Keiki's lab could help you with the implementation of that.
[15:35:09Z] When we were looking at the output of our code I remember scrolling through lots of similar-but-not-the-same rules to see whether it had got close. Having a higher-level description of that would have meant we could have used larger populations, I guess.
[15:36:57Z] Though the range of possible solutions to what we were asking it to do were pretty small.
[15:38:14Z] yeah the idea is to replace the scrolling with something better
[15:39:11Z] it could start simple but be extended in lots of ways e.g. computing equivalence between rules would be a nice extra feature
[15:39:22Z] Could the tool capture something that the user learned about the data, parameters, etc, and feed that back into the experiment?
[15:39:39Z] meta-learning?
[15:40:18Z] Keiki’s people would find it useful. if you came to japan that could be your official project, if you wanted.
[15:41:00Z] Well I mean more like the user learns something new about the thing they are trying to model, like the rule language or domain has a new equivalence that the XCS didn't take avantage of, or something.
[15:41:32Z] Yeah, I'm interested in tools that let you kind of elaborate your understanding of some domain and use that to simplify or extend the learning process.
[15:42:40Z] XCS problems are typically quite small (in features), but it might help simplify larger problems enough for them to be more tractable.
[15:42:54Z] I guess I'm mixing up the XCS and the exhaustive search now.
[15:43:00Z] I think it could although I don’t have an example in mind. I guess the idea is to make the rule set less of a black box. I’m motivated by getting the box to work better, but it would help with domain understanding too.
[15:43:43Z] I think that's a general problem with the state of data mining at the moment. Domain understanding is just one solution.
[15:44:10Z] they are typically small as it doesn’t scale well
[21:12:20Z] Welcome Joey! What are your spare-time projects at the moment?
[21:13:10Z] (if any)

All timestamps are in UTC. If you want to join in, channel information is available. This page is generated from raw logfiles dlab-2017-10-25-164609.txt, dlab-2017-10-28-155634.txt, dlab-2017-10-28-175235.txt.