Author Topic: AI using Probing attacks  (Read 907 times)

Offline DrFranknfurter

  • Full Member
  • ***
  • Posts: 116
AI using Probing attacks
« on: May 06, 2014, 04:09:39 pm »
I read in the thread about threat working differently this quote from Keith:
Quote
So there's the waves as probing attacks, which (so long as you have an AI warp gate somewhere on your border) go directly into the teeth of your defenses.  Nobody on either side expects those to win a game, they're just there to probe your defenses at different points with different unit combinations and maybe cause a ruckus if you've made a mistake or it lands in the middle of general chaos, etc.

I was just wondering about AI probing attacks, adapting and learning to counter you. So here is my idea for expanding that probing attack concept. Could the success of standard waves be measured? (It's critical to this entire idea) Then used to adapt to your ability to defend. Perhaps as a future AI type: Adaptive

Success:
So how do you measure success? I think damage alone would favour bombers blasting away at high health tractors, survival time (of all or just a few 'marked ships') would favour tanks, armoured ships or sneaky cloaked ships evading defences, kill count would favour railpods/snipers/glass cannons... So why not use all of them?
Success = Wave survival time (0-5min, max 1k points) + Damage dealt (0-?? no idea really??, max 1k points) + kills (0-1000kills, max 1k points) + AIP bonus (AIP*2, max 1k points) + Difficulty bonus (0-10, Diff^3 max 1k points)
Success magnitude (0-5ish, points/1000, rounded down) i.e. Always have 1+point Success every wave at diff 10
Failure magnitude (0-5ish, points as above, all except diff+AIP are inversed) ~3+point failure at diff 10, per wave

Adaptations: Randomly choose a different adaptation after each wave is analysed (at 5min after wave)
1. Change wave size/budget (wave interval)
If Success>failure, attack faster with slightly smaller budget (Points converted to budget adjustment)
If failure>success, attack slower with slightly larger budget

2. Change cost of fleetships/starships (reduced costs = more of them)
If Success, future waves with same type have slightly reduced costs (0.1xSuccess%, up to a max of AIP% increase in the effective budget for that ship type)

3. Change weighting of ships.
If Success>failure, pick ships more often (up to 0.5*Difficulty, so max of 5x more often at diff 10)
If failure>success, reduce weighting slightly (Down to 2/Difficulty, min of 1/5 as often at diff 10) OR,
If failure>success, cycle to next ship type

4. Suppliment next waves with a successful ship type
If Success of a ship type > 1, then (Difficulty*Number of successes, max 100)% chance to modify next  wave, 1/10 that chance to modify every wave. Adds (AIP/20+Difficulty+No. Successes)% bonus of all successful ship types to next wave. (max bonus being AIP% of normal wave size)

5. Reinforcement pulse of successful ship type
As above, chance to add to planet reinforcements or strategic reserve either in next pulse or all future pulses.

Failed probings?
Each failed probe point (normally every wave) can add:

+1 effective AIP only used for unlocking new ship types, +0.5 at diff7. (scary)
+1% Exo charge (could allow for continually synchronising, rapid exos, probably evil)
+10% wave budget for next wave (potentially problematic but transient)
+5% ships in next wave increase in mk. (potentially problematic but transient)

Thoughts?

Offline keith.lamothe

  • Arcen Games Staff
  • Arcen Staff
  • Zenith Council Member Mark III
  • *****
  • Posts: 19,505
Re: AI using Probing attacks
« Reply #1 on: May 06, 2014, 05:42:34 pm »
By probing I didn't mean that it recorded the results or whatever, just that it keeps trying different things against different places, which tends to find chinks in your armor if they're there to be found.

"Remembering" that something worked well last time is actually not generally advantageous.  Presumably, if the wave worked well enough to be worth remembering then one of two things has happened:

1) The player is dead.

or

2) The player is not dead, and is now angry, and is very likely to ensure that a repeat performance will meet with very different results :)

And often a successful wave is successful not because it singlehandedly overpowered human defenses but because of other chaos that was going on.  Chaos which is probably not happening in a sufficiently similar fashion in the future.

As far as having failures make the next wave bigger, that gets into the territory of penalizing the player for playing well, which I would like to avoid.


Anyway, all worth thinking about but ultimately there are good reasons it doesn't actually try to learn.  Its only teacher would be the enemy :)
Have ideas or bug reports for one of our games? Mantis for Suggestions and Bug Reports. Thanks for helping to make our games better!

Offline DrFranknfurter

  • Full Member
  • ***
  • Posts: 116
Re: AI using Probing attacks
« Reply #2 on: May 06, 2014, 06:03:18 pm »
I see the counter-argument, though it would be a way to spice up normal waves significantly in games without any exo sources (but less harshly than beach-head) and add a bit of thought and possible meta-gaming depending on the exact mechanism.
e.g. Don't kill fighter waves too quickly (make it count as a success), or avoid using enclaves (lots of drones dying, not quite as bad as with a vengence generator but gives a skewed kill-count)

There are already AI types that punish 'success', launching counter waves, blowing up etc. And people like pain... it seems. Ah, well you've read it anyway so that counts as a success. +1 point. Also +1 point of failure but +1 step towards unlocking a new AI-type

Offline Lancefighter

  • Core Member
  • *****
  • Posts: 2,440
Re: AI using Probing attacks
« Reply #3 on: May 06, 2014, 06:12:59 pm »
You should enable sharks.. Those include.. "learning" from successful waves. (also!!fun!!)
Ideas? Suggestions? Concerns? Bugs to be squashed? Report them on the Mantis Bugtracker!

Author of the Dyson Project and the Spire Gambit

Offline keith.lamothe

  • Arcen Games Staff
  • Arcen Staff
  • Zenith Council Member Mark III
  • *****
  • Posts: 19,505
Re: AI using Probing attacks
« Reply #4 on: May 06, 2014, 06:14:13 pm »
You should enable sharks.. Those include.. "learning" from successful waves. (also!!fun!!)
It does generally involve being chased by large enraged animals who are also on fire.
Have ideas or bug reports for one of our games? Mantis for Suggestions and Bug Reports. Thanks for helping to make our games better!

Offline DrFranknfurter

  • Full Member
  • ***
  • Posts: 116
Re: AI using Probing attacks
« Reply #5 on: May 06, 2014, 06:35:54 pm »
Oh, so many interesting ways to die. I do quite like concept of those options that modify the standard gameplay (beach heads, munitions/armour boosting support AI-type, any AI type that spawns special structures). Though I'm struggling on difficulty 8 so it'll be a while before I get around to activating each option at each intensity and longer still before I make my way through every AI type. The Shark plot (B is it? the one with the nasty waves not the AIP increase, more a slow bleed into the water that one) that sounds... fun (no sarcasm, just challenging, A+B could be a pain though... or never noticed at all, depending on how it goes)

Still, some AI-types require entirely new strategies to play against or make something that should be easy hard. Normal waves get left behind a little... they don't have the punch of an exo, are often quite far apart and yet... they happen so often that any big change to them you'd notice right from the word go. It feels like a decent bang for coding buck to me.

There are so many cool and interesting mechanics that hide away until they leap out at you... but I think you'd notice and feel worried if after a wave you got a message
"The AI has learned from its last attack...
"...wave budget increase to level 2"
"...wave speed increased to level 2"
"...Poan is vulnerable"