| |
| |
| |
|
Comments:
<0> Sotek, hmm..
<0> Does it have to check in performance?
<1> Andares: Hmm? Not ... certian what you mean.
<1> ...certain. damnit.
<2> http://www.deadbeefbabe.org/paste/1031
<2> this is an output of 20x20 map
<2> with 50% random cell costs of 10-99
<2> and the other 50% are just cost of 10
<2> the closed list is too big in my oppinion
<2> it checks way too many nodes to find a the shortest path
<2> i think that's where my bottle neck is
<1> citrus: Agreed. That's why I was suggesting you might want a more accurate heuristic.
<2> hmm
<1> is your heuristic right now manhatten*10?
<2> no
<2> i tried with manhatten*10 and it overestemates sometimes
<2> and doesn't give the shortest path
<1> oh yeah, of course it does.
<1> since it can go diagonally.
<2> yes
<1> what you want is max(x,y)*10
<2> when it goes diagonaly, i do compute it as a more cost
<1> for something simple.
<1> ah, hmm.
<2> hmm
<1> how so? 1.5?
<1> or what?
<2> i take square root of the 2*the distance to that cell
<2> that way it looks like a pythegorian triangle, with sides A, B and i comput C by Sqrt(A^2+B^2)
<2> so instead of say ... 10, it becomes 14
<1> hmm.
<2> but
<2> what do you say about max(x,y)*10?
<1> okay, 10*max(x,y) + 4*min(x,y) should be decently accurate and never overestimate.
<2> x, y of the current node?
<2> or distance to the goal
<1> x and y being the difference.
<2> ah
<2> ok
<1> modified manhatten, basically.
<2> hmm
<1> the goal of an A* heuristic is threefold: 1) you want it to produce numbers as high as possible, while 2) never overeestimating, and 3) being fast to calculate.
<2> yes
<2> ok
<2> i like it, it's alittle better
<2> but still too many nodes on closed list, i think
<2> http://www.deadbeefbabe.org/paste/1034
<2> takes 4.458 seconds to do 100x100 map with random cost values
<1> hmm. I dunno, that looks decent. What if you start from the center and go to a corner?
<2> ok
<2> my map isn't always pure random
<2> there is alot of "same terrain"
<2> your heuristic is very good
<2> what my original manhatten distance used to solve in 4.5 seconds
<2> with your heuristic it's .3 seconds
<2> i'm not sure how corrent the path is yet, because the map is randomly generated every time
<2> heh
<1> hee. well, mine does get a bit closer. But it's A* - you can tell for certain if it's correct or not, simply by determining if the heuristic is capable of overestimating.
<2> it never follows straight line
<2> it always takes the NE or SE or NW or SW directions
<2> yea
<1> Hmm. Never? Not even if it's going horizontally?
<2> yes
<2> it just goes like a snake
<1> Hmm. Interesting.
<2> yes indeed
<2> but, it's hella fast
<1> Well, if it doesn't get shorter by going straight, then not that surprising. If it does, then that's a problem. hmm.
<2> http://www.deadbeefbabe.org/paste/1035
<2> this is what i used for heuristic
<1> And, well. Mine is a bit more accurate than manhatten - note that if 10 isn't the min cost, mine could be wrong, since I did ***ume that.
<2> hmm
<1> (just change 10 and 4 to 1 and .4, or whatever, if so)
<2> http://www.deadbeefbabe.org/paste/1036
<2> take a look
<2> even map
<2> everything is 10
<1> ... oh wait. Diagonals don't add extra cost.
<1> Length: 20 Total Cost: 190 <-- pretty clearly, then...
<1> so my heuristic is actually wrong because it /can/ overestimate. Just 10*max(x,y) would be better.
<0> Why don't diagonals add extra cost?
<1> (or change it so that they DO cost 1.4 times, and stay with mine...)
<2> hmm
<2> the path looks alot better
<2> the number of additional visited nodes increased as well
<2> in a given map, between point A<x1,y1> and B<x2,y2> it visites about 80% of all nodes in that block of graph
<2> that is between x2,x1 and y2,y1
<2> this isn't good still
<1> It's probably going to have to.
<1> I do expect that if you have "walls" or the like in there, though, it'll not try the nodes "shadowed" by them...
<0> Hi.
<3> hi
<4> hi
<3> Hey Doctor Chandra.
<3> i looked at a AANN, and it seems the more they added linear and nonlinear hidden node, the mode the results looked liek the orginal test image.
<4> sure but the interesting part is the middle layers
<4> the signal they form
<4> those are interpretations
<4> 'understanding' if you will
<4> concepts in the images it is exposed to
<3> I gathered that the idden layers create the differences between the nodes abosolute value's inreased double fold and iterated over ever start and end node. Is it anything more than this or I am missing it complety?
<4> I'm not aware of doubling of any values
<4> it's a normal neural net you train with input as both input and target output. the unit function. the interesting part then shows up as the output values of the middle layers since they express some general concepts on the dataset which it was exposed to
<4> it tries to learn a function that seemingly does nothing to the signal
<4> the same could be said for compression/decompression yet the middle result is of great interest. the same is the case here
<3> the sources for bzip2 would make an exxellent music compator engine. 'Fetch me music that sounds like this + that"
<5> anyone familiar with constraint programming here?
<6> Hi.
<6> DanF_DrC, you there?
<6> Hi petroleum.
<6> What is the best way to feed an ANN strings of text?
<6> Do you stream them somehow or organize 30 or so inputs, ***uming 30 chars max
<6> ?
<7> <Luke1704@ef> hello
<7> <Luke1704@ef> anybody here?
<7> <Luke1704@ef> I am searching someone how may help me in an C++ problem
<7> <Luke1704@ef> I have an implementation problem with a template
<6> Luke1704, you still there?
Return to
#ai
or
Go to some related
logs:
#fedora
alias out of office postfix courier
codec umbutu
dot-qmail-exists.pl
fedora w32codec.rpm
hwu54g+ubuntu
centos uninstall cpanel
#php
pykota
#perl
|
|