This bug can be game breaking: Motion Prediction

[camellia0x++]

Motion prediction is very bad, it allows your own units to be hit. Please make it either check for walls, or at least your own units. A clever person can use this to have your own artillery destroy your forts/walls/hardpoints. Even bots do it randomly.

[aubergine]

It's just part of the game - as you research more you get better target prediction and eventually things like search and destroy missiles that home in on their targets. The trick is to place defences wisely, and if your walls are too weak research more wall upgrades (they quickly become immune to your own mortar fire after a few upgrades for example).

[Emdek]

aubergine is right, however option to disable friendly fire could be interesting.

[NoQ]

A clever person can use this to have your own artillery destroy your forts/walls/hardpoints.

Well, that's a ... war ... isn't it? What's wrong with being clever and using tricks to win?

The whole friendly fire is pretty messed up btw. For example, flamers don't do friendly fire at all, even though there is a lot of flame burning around; but if you make them do friendly damage, they will instantly burn themselves to death.

[camellia0x++]

It's just part of the game - as you research more you get better target prediction and eventually things like search and destroy missiles that home in on their targets. The trick is to place defences wisely, and if your walls are too weak research more wall upgrades (they quickly become immune to your own mortar fire after a few upgrades for example).

The issue is that (if this were real,) it would be very simple to check with HQ if there is a structure at that location, these aren't misses, it's just stupidity.

Also, bots may continue "walking" into walls, at which point, they are treated as moving, even though they are still, so the motion prediction is buggy, using computer logic (the vector variables?) instead of physics (it isn't actually going anywhere while walking into the walls.)

[NoQ]

It's a hard desicion wether a howitzer should fire risking damage own units or shouldn't fire at all leaving enemy alive. I wouldn't like it to be automated.

Also, bots may continue "walking" into walls, at which point, they are treated as moving, even though they are still, so the motion prediction is buggy, using computer logic (the vector variables?) instead of physics (it isn't actually going anywhere while walking into the walls.)

Yep, that's the real problem (:

[aubergine]

I quite like the fact that missile prediction works the way it does, it makes the game more interesting. If everything was 100% accurate and missiles never missed, it would be boring IMHO.

Also, remember that the missile was fired before a droid ran in to a wall and stopped. So the missile predicted where the droid would be at the time the missile was fired.

[Lord Apocalypse]

The issue is that (if this were real,) it would be very simple to check with HQ if there is a structure at that location, these aren't misses, it's just stupidity.

You also have to figure the AI into things. In a real war you wouldn't want to hit a wall 10ft in front of you or you and your gun would blow up. The game AI on the other hand may know a friendly is in front of it but may not know that the friendly is a wall that is much higher than a regular unit. Also, most artillery shouldn't be able to hit something so close anyway.

Also, bots may continue "walking" into walls, at which point, they are treated as moving, even though they are still, so the motion prediction is buggy, using computer logic (the vector variables?) instead of physics (it isn't actually going anywhere while walking into the walls.)

This isn't an issue with the physics but more of one of AI pathfinding and the fact that the unit AI should be able to say "Hey, there's a wall here I better look for another way to go". Also, collision detection should also tell us that unit A has hit a wall or come inside the bounding box of another object and one of the objects needs to go somewhere else. Game logic could then tell us if the unit is still trying to push over the wall (not moving) or chosen a new path (still moving)

[camellia0x++]

I quite like the fact that missile prediction works the way it does, it makes the game more interesting. If everything was 100% accurate and missiles never missed, it would be boring IMHO.

Also, remember that the missile was fired before a droid ran in to a wall and stopped. So the missile predicted where the droid would be at the time the missile was fired.

Problem is they don't stop walking after they run into the wall. They keep trying to cross it.

It's a hard desicion wether a howitzer should fire risking damage own units or shouldn't fire at all leaving enemy alive. I wouldn't like it to be automated.

Maybe have new right click options for them?

[Jorzi]

I wouldn't call it game breaking, really. Many games have problems with the stupidities of AI:s, it's something you learn to play with (i'm not trying to defend the stupidity of the AI, a smart well-working AI is really valuable). After a while you start seeing how the AI "thinks" and once you know its strengths and weaknesses you can adapt your playstyle to it.
That said, I welcome any improvements wholeheartedly.

[aubergine]

There's some plans to create long-term persistence objects that AIs can use to store data from one game to the next. This will enable scripters to start experimenting with AIs that learn about how the human player "thinks" and know the human's strengths and weaknesses (note: It will take a lot of time and trial and error to get that sort of thing working reliably)

[Emdek]

aubergine, ah, now I see good use case for storing persistent (independent from individual games) data, good idea.
I think that it might be also needed to save some aspects specific to used map (like what players does first etc.).

[aubergine]

Exactly In effect, an AI should be able to store data about it's engagements against a human player at various stages of the game for a specific map. So it might know that sending a load of MG tanks to a specific area at one point in the game works well against this player, whereas 10 minutes later it turns out to be a disaster.

It's going to take quite a lot of effort to find a reliable but fast way of collecting, storing and then processing the data, so realistically it's probably going to be 2013 before we start to see any major advances in this approach.

[Lord Apocalypse]

@aubergine Try a neural net based AI, or at least using a neural net for the AIs backend. There are references all over the net as well as a few (at least 1) on sourceforge I think.

Its either that or expanding the games AI scripting and I don't know how far you can take that until it breaks or becomes a bit too buggy.

[aubergine]

It just so happens I've tried developing a neural net in Javascript before. It didn't turn out too well. There's actually some stuff in the more recent editions of ECMA-262 that will help make things more manageable, but they're not yet implemented in the QtScript environment that Warzone uses.

A bunch of the things that would be useful to me are summarised in this blog post by Talia Bale: http://taliabale.tumblr.com/post/104325 ... proved-for

I was actually testing a load of that stuff in the WZ JS environment last night hence having the link to hand. Unfortunately, it seems that QtScript isn't quite as complete an ECMA-262 implementation as I had first thought (although it's got loads of great stuff in there that's already making my life easier).

Things like deconstructing, proxies, weak maps, array comprehension and so on, remove a lot of the barriers I ran in to last time I tried implementing neural nets, primarily because they make the complex coding more manageable and also let me get access to some lower level stuff that would be too slow/clumsy to do with higher level functions.