[Ohrrpgce] SVN: james/8760 Add an A* pathfinding for maps. NPCs can be set to chase the hero around

Ralph Versteegen teeemcee at gmail.com
Sun May 28 20:40:00 PDT 2017

On 29 May 2017 at 02:52, James Paige <Bob at hamsterrepublic.com> wrote:

> On Sat, May 27, 2017 at 7:49 AM, Ralph Versteegen <teeemcee at gmail.com>
> wrote:
>> On 27 May 2017 at 11:35, <subversion at hamsterrepublic.com> wrote:
>>> james
>>> 2017-05-26 16:35:19 -0700 (Fri, 26 May 2017)
>>> 346
>>> Add an A* pathfinding for maps. NPCs can be set to chase the hero around
>>> obstacles.
>>> Currently only supports wallmap checking, I will be adding zone-checking
>>> and NPC-collision checking soon.
>>> You can test this already, but the NPC movement type is marked
>>> "EXPERIMENTAL" because the behavior will definitely change a bit before I
>>> am done with it.
>>> ---
>>> U   wip/SConscript
>>> U   wip/customsubs.rbas
>>> U   wip/game.bas
>>> A   wip/pathfinding.bas
>>> A   wip/pathfinding.bi
>>> U   wip/util.bas
>>> U   wip/util.bi
>>> U   wip/walkabouts.bas
>>> U   wip/walkabouts.bi
>> Excellent!
>> I was going to remind you about one-way walls, but looking quickly at the
>> code, it looks like they would just work automatically.
>> How will you handle tiles blocked by other NPCs? Just give them a high
>> cost? I don't know a good solution for this. If the cost isn't high enough,
>> then the NPC won't go a different way while there's an NPC in the way if
>> the other way is an extra X files.
>> Really, any stationary NPC should be treated more like a wall. But so
>> should NPCs that are just stuck in a corridor.
>> Maaaaybe this could be handled by giving each NPC a dynamic path cost.
>> For every tick that another NPC is trying to get past it (its path goes
>> straight through the blocking NPC), its cost increases. Finally the pathing
>> NPC gets so sick of waiting that it goes another way. This could happen
>> before the pathing NPC ever reaches the blocking NPC, since the cost can be
>> updated for each tick that the pathing NPC moves. I don't know what the
>> condition would be for decreasing the dynamic cost. Maybe something like
>> when the NPC moves off someone's path. If an NPC gains a really high cost,
>> then computed paths across the map with fluctuate wildly when the high-cost
>> NPC steps in and out of a corridor.
> [re-sending this reply too. I think the server hosting the mailing list
> had a full-disk last night, but dreamhost has fixed it by today]
> I'll definitely treat NPCs with a non-moving movement type as walls, and
> I'll probably just go with giving moving NPCs a higher pathing cost but
> still treating them as if they are permeable. I don't think there is a
> perfect solution for this, so I am going to err on the side of a simple
> solution, and give it a bunch of testing. I was also thinking of giving
> them a higher cost when they are close (thus having less time to get out of
> the way) and a lower cost if they are far away.

But there are many reasons why an NPC not set to Standing Still won't move.
So I think it is better to use a counter for how long it's been since the
NPC moved rather than rely on the move type. You could try to add logic to
determine all the reasons an NPC isn't going to move in future, too.

NPCs could also be set to being pushy, meaning they will swap places with
another NPC they're trying to get past, provided that both NPCs would be
otherwise able to make that move. This is sometimes used in roguelikes so
that a boss enemy pushes its way to the front of a crowd instead of
becoming stuck.

> Testing will show whether I need to add a bunch of configuration options
> for this or not.
>> Have you tried running this on larger maps? In particular when you place
>> some NPC that can't reach the hero? I image that in that case the NPC would
>> try to repath every tick. We're going to have to stop that from happening.
>> Maybe delay for half a second before trying again. It's also possible to
>> cancel further pathing attempts until either the map changes or the hero
>> crosses a one-way wall (approximation: when the hero/target changes).
> I haven't done any profiling or stress-testing of this yet, but I plan to.
> A cooldown time before an failed-path NPC tries to path again is a great
> idea. Not sure if I will make it configurable or just pick a reasonable
> default.
> Right now I expect that most of my pathing time is wasted in re-allocating
> a new pathfinder object each time an NPC wants to use it.
> I hope to have just one pathfinder object that all the NPCs on the map can
> share.
> I read about some optimizations that are possible by using OPENED/CLOSED
> values that increment +2 on each new call rather than being constants, so I
> can just allocate the open/closed status array once per map, and only have
> to flush it clean once every billion calls.
> I don't plan on actually trying that one until after I have some profiling
> tests of some kind set up so I can know how much it helps.

I think you're looking at the wrong thing. On a large map, when computing
very short paths, the array allocation time would dominate, but it's going
to be a negligibly small part of the worst case running time, when there is
no path.
Note that the line "status as AStarNodeStatus = AStarNodeStatus.EMPTY"
causes FB to create a constructor for the type, which slows down redimming
and deleting the array. With -gen gcc I found that gcc is able to optimise
it, so the slowdown is only 50%. With -gen gas, the slow down is enormous,
over 200x slower. Taking 1.5ms on a map with 100k tiles (the max allowed
map size).
Without the constructor, the redim only takes 8us on my machine for a 100k
map (regardless of gen gcc/gas). It could even be sped up more, by reducing
the size of AStarNode.

(I also noticed that without the constructor, compiling for x86_64 made the
redimming 10x-80x slower on my machine, even after I made sure the UDT size
was the same and if I use gen gcc both ways. It appears to be a libc
problem: all the time is spent in calloc, but the codepath used on x86_64
seems to be very slow. Probably it's inappropriate for my CPU; it's using
some fancy new instructions I've never heard of.)

(BTW, I would like to switch all releases and nightly builds to gen gcc,
because the binaries are faster, smaller, and bugs that currently only
affect -gen gcc will be found faster.
(Also, the currently Windows nightly builds are huge. The .exes are aronud
twice the size that they used to be before switching to the new build
machine. I haven't looked into why yet.))

Anyway, looking at the code (haven't profiled), I think the real
bottlenecks are check_wall_edges, function call overhead including passing
around AStarNodes, searching for the best node, and v_append and v_remove.
Of course, the later few will be to blame for all runtime the worst case.
getnode is called excessively. I never optimised vectors to avoid calling
realloc every time that the vector is resized. I've been meaning to.

Oh, I notice that guess_cost_after_node doesn't work on wrapping maps.

>> I notice that how you determine the 'best' node of the openlist is very
>> inefficient. Instead, you should keep openlist sorted so you can pop off
>> the end (popping off the front is slow). You can use v_sort to sort the
>> list, but first you need to define a compare function. See vector.bas for
>> examples.
>> Using v_sort on an almost-sorted array is still less efficient than it
>> could be (linear instead of logarithm time). I have some C code for a
>> priority queue, so I could add a couple functions for using a vector as a
>> priority queue, and we can switch to that later.
> I saw your other reply about v_sort
> There is another reason I like open-list sorting that is unrelated to
> performance. I could have my sort-order resolve ties in a well-documented
> and minimally biased way. Right now I just favor the first closest step
> found, which has some directional biases based on the fact that I loop
> north/south/east/west around each next node.

So how would you break ties?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.motherhamster.org/pipermail/ohrrpgce-motherhamster.org/attachments/20170529/79c9c04c/attachment.htm>

More information about the Ohrrpgce mailing list