New Bot AI Progress

[jitspoe]

Oh, what a tangled web we weave!

I think the dynamic waypoints I've done are about as good as can be expected.  Looking at the results, I can't help but think, "if this were a navmesh, it could guarantee 100% coverage and not have to have nodes everywhere with the potential of missed paths."

The disadvantage to nodes, is you never really know when you have enough.  With a navmesh, you can say, "I'm on the navmesh" and be done with it, or "I'm not on the navmesh -- need to generate another poly."

I think I'm going to continue with waypoints for now since I've gone this far.  Ultimately, the bots will hopefully be following the other path system I'm working on to allow strafe jumping, and this will just be there as backup/beginner level bots.

I have A* pathfinding implemented.  Now I just need to make the bots follow some paths and give them some goals.

[Clipz]

Will bots be smart enough to drop a brass barrel for a steel barrel? To drop a spider for a carbine or auto cocker?

[not_payl_obviously]

Looking at screenshot I think there are too many paths, since bots would mostly use just small part of this.
Maybe there should be color indicating what kind of path it is.

Also I see one path being impossible to do upwards, is detecting this implemented?
Jitspoe, what about some pretest soon, so we have some time to play with new bots and test what is most imporant to be done? I'm not sure how complete is this new AI in basic matters, but I think pretty much anything will be better than old bots.

Will bots be smart enough to drop a brass barrel for a steel barrel? To drop a spider for a carbine or auto cocker?

If someone implements this. Jitspoe said he is going to make open bot library, so even you can implement this.

[BASEBALLDUDE]

If a bot somehow gets knocked off-track, will it be able to find its way back on track?

[jitspoe]

Will bots be smart enough to drop a brass barrel for a steel barrel? To drop a spider for a carbine or auto cocker?

Not currently, but maybe... trying to decide if it's worth putting any effort into items since about the only maps anybody plays anymore spawn you with equipment.

Looking at screenshot I think there are too many paths, since bots would mostly use just small part of this.

Mostly, but...

If a bot somehow gets knocked off-track, will it be able to find its way back on track?

The more waypoints there are, the less likely bots are to get stuck in places.  Also, when a flag might get dropped in some weird spot or a bot needs to track down a player hiding somewhere.  Anywhere players go, nodes are added.

[not_payl_obviously]

The more waypoints there are, the less likely bots are to get stuck in places.  Also, when a flag might get dropped in some weird spot or a bot needs to track down a player hiding somewhere.  Anywhere players go, nodes are added.

True, but still maybe mark somehow paths that are traveled most often? You said you implemented path finding, how bots decide what path of map to use? Will this vary? What about CPU usage?
With paths that are traveled mostly we could mark paths and it would reduce a lot of complexity in pathfinding.
Still thing you should focus on the most is some kind of strafejumping. It is basic in this game, then we can add learning from player some more complex jumps.

[jitspoe]

True, but still maybe mark somehow paths that are traveled most often? You said you implemented path finding, how bots decide what path of map to use? Will this vary? What about CPU usage?
With paths that are traveled mostly we could mark paths and it would reduce a lot of complexity in pathfinding.
Still thing you should focus on the most is some kind of strafejumping. It is basic in this game, then we can add learning from player some more complex jumps.

For the moment, it just uses the shortest path.  I want to do something based on probability, like Markov chains.  This would make the bots behave, statistically, like players, though they wouldn't necessarily make the best decisions for a given scenario.  They might give the illusion of being smart by using different paths that players use in different scenarios instead of taking the same paths every time (ex: what's the probability of going to each node connected to the waypoint, after grabbing the flag).

[Union]

I hope you succeed with the new AI.
I hate the bots we have now because it is like they have aimbot, and they don't do much.

[jitspoe]

I made a little video to show how some of the waypoint and path finding stuff works.  Uploaded to Vimeo and YouTube (pick your poison):

https://vimeo.com/116576240
https://www.youtube.com/watch?v=jphSKHwGjOw

[Bundy]

This looks great!

[Rick]

If this does not have audio commentary.... I'mma lose it.

Edit:   

[FusSioN]

Jitspoe's voice tho. Nice work!

[T3RR0R15T]

What's the difference between the white and purple lines?

[jitspoe]

What's the difference between the white and purple lines?

They start out purple and get lighter the more frequently they are updated.  I think I was going to do some averaging for the waypoints to try to center them in the most frequently passed areas, but ran into a lot of edge cases, so I just ended up relocating close way points to exact player location to guarantee they were valid.

[shockk]

The whole bot problem with jumping up the boxes in that video... Isn't there just like a value you can increase in your code that makes them attract to each way point more accurately?

Like 5 as a value for example, they'll be near the way point then move on, but if you set the value to 20 you'll see them get way closer to each way point?

Hard to explain and I don't know this stuff but thought I'd mention it

[jitspoe]

It's easy to tweak how close tho bot should be to waypoints before considering them reached, but if it's too strict, the bots will often "miss" waypoints (especially on slopes) and have to double back to hit them.  If you're suggesting having a value on a per-waypoint basis, that could also be done, but it would involve introducing a new system for determining what that threshold should be at each waypoint, or have the bots somehow "learn" which waypoints they struggle on and tighten those thresholds.

[shockk]

Yeah I don't know much about it at all, but if the bots could recognize an "attempt" at reaching a waypoint (with a strict accuracy) maybe they could then register whether the waypoint was reached or not, and be programmed to tighten the parameter (raise the accuracy) on the next try before the box is reached and the waypoint attempt was verified as successful by the bot.
It would be just like someone trying to jump up a ramp, realizing it wasn't high/close enough, then re-attempting before verifying it as reached.

[shockk]

And then moving on to the next waypoint of course. But basically I was thinking something like adding an "attempt to reach waypoint" sort of process, then a "reached or not completed" follow up. Then they could readjust based on what values are increased automatically for the bots or whatever is programmed for them to changed

[Cameron]

In the case of those boxes, the objective is to come to you, in which case, it should only pass through each waypoint once.  So say it passes through a waypoint twice, would it be possible to tweak how close it should get to the next few waypoints (or up until that last waypoint it reached before and then fell), just to help it get past that bit?  I assume there would have to be some time window for this to apply though.

[TRION]

So to make the bots better we will have to let them play on a map for hours?
Will they have personalities like lagger,newbie,pro,just a good jumper and "doesn't even know he is now in red"?