Ah, good question, no I haven't. I'm basing that comment on a vague memory of a graph showing query time vs preprocessing time for lots of algorithms. Looking back at the graph it seems ch with arc flags (chase) slightly improves query time, but for alt they only have results on non fully contracted networks (core-alt). Probably a memory issue.
Sorry if I misled - I haven't actually tried combining ch with alt as it turned out to be more than enough on its own. My point was just that ch is reported to synergise well with other methods, which makes it likely to be a good investment.
Interesting, we (GraphHopper) tested CH+ALT and saw also no improvements, also for CH with normal A*, although it is reported in papers that it should be faster. The idea about non-fully contracted networks is really nice and could be worth a try - thanks!
Wow, that seems counter-intuitive. I suppose it must be due to the computation of the heuristic, combined with how few relaxations are needed in CH.
GraphHopper is really nice work by the way. I believe you are still working on implementing stall-on-demand? Will be very interesting to see how much difference that makes on long-distance queries.
I guess it is the overhead of the heuristic that is too much compared with the saved visited nodes, but this is counter-intuitive for me as well and if I have more time will investigate this again.
> GraphHopper is really nice work by the way.
Thanks!
> I believe you are still working on implementing stall-on-demand?
Sorry if I misled - I haven't actually tried combining ch with alt as it turned out to be more than enough on its own. My point was just that ch is reported to synergise well with other methods, which makes it likely to be a good investment.