PolyCast episode 386: Machine Learning

[megabearsfan]

The three hundred-and-eighty-fifth episode of PolyCast, “Machine Learning“ is now available for streaming on polycast.civfanatics.com. This episode features regular co-hosts
Stephanie “Makahlua”, Phil "TheMeInTeam", and Jason “MegaBearsFan”. Topics for this episode include:

Forum Talk
- 00h01m29s | Civ VII could Make It's A.I. Smarter and How fixing the A.I. should be easy
- 00h37m41s | Firaxis' "Impossible Save" is not-so-impossible.
- 00h43m05s | Outbreak of mange
Senate
- 00h47m46s | Why did the April patch buff tanks?


PolyCast is a bi-weekly audio production recording live every other Saturday at 12 Noon US Eastern Time throughout the year, in an ongoing effort to give the Civilization community an interactive voice; sibling show ModCast focuses on Civ modding.

I apologize for the audio equalization of this episode not being so great. I may have missed a step during my usual editing and normalization process, and didn't have time to go back and correct it.

 

[Deggial]

Guys, you are missing one crucial point in your AI discussion, when comparing Civ with Starcraft (and Dota, I guess, but I’ve never played this game):
As far as I know, the Starcraft AI is learning to play on one very specific map. Only ONE map! It learns, when to expand, where to expand, which routes to take when attacking, where to defend. Furthermore, the terrain is actually very simple in Starcraft: high-ground, low-ground and walls.

In 4X games such as Civ, the terrain is usually procedurally generated and different for each game. And the created terrain itself is highly variable, with a large impact of terrain type on the unit behavior (speed, attack/defense values ect.)

This is a huge difference in the demands for a self-learning AI!

Edit: city development and all related tasks are probably way easier learn for an AI than strategic planning and tactical unit movement, though.

 

[frutiemax]

Also, the goal of alphastar is to beat professionals at starcraft, while in Civ6 we're only talking about some sort of basic intelligence.

 

[megabearsfan]

Guys, you are missing one crucial point in your AI discussion, when comparing Civ with Starcraft (and Dota, I guess, but I’ve never played this game):
As far as I know, the Starcraft AI is learning to play on one very specific map. Only ONE map! It learns, when to expand, where to expand, which routes to take when attacking, where to defend. Furthermore, the terrain is actually very simple in Starcraft: high-ground, low-ground and walls.

In 4X games such as Civ, the terrain is usually procedurally generated and different for each game. And the created terrain itself is highly variable, with a large impact of terrain type on the unit behavior (speed, attack/defense values ect.)

This is a huge difference in the demands for a self-learning AI!

Edit: city development and all related tasks are probably way easier learn for an AI than strategic planning and tactical unit movement, though.

That is an excellent point. I wonder if that would change TheMeInTeam's opinion of how well it would work for Civ? Hopefully, in the next episode, I'll remember to bring that up.

 

[Aristos]

Behavior trees, the AI paradigm selected for Civ 6, is already obsolete, but more importantly, it is not fit for complex behavior (i.e. complex turn based systems). Ironically, as far as I remember, the AI designer for Civ 6 was ex-Starcraft, where he supposedly implemented BTs and brought that experience to civ 6. Big mistake.

Utility AI is the best system for TBS. Ironically again, Civ 5 scratched the surface and implemented a barebones Utility AI framework (they called it weight-based, and you can see it clearly in the code). Even if barebones, it is easy to see how superior vanilla civ 5 AI is to Civ6 BT AI. Gazebo and Ilteroi perfected the Utility AI system in Civ 5, and we got the infinitely superior AI in Vox.

There is no need for ML in civ yet. Utility AI is more than enough for the time being, and still provides powerful opponents while keeping turn times under control.

See here:

https://www.gamasutra.com/blogs/Jak...88/Are_Behavior_Trees_a_Thing_of_the_Past.php

 

[Gedemon]

Behavior trees, the AI paradigm selected for Civ 6, is already obsolete, but more importantly, it is not fit for complex behavior (i.e. complex turn based systems). Ironically, as far as I remember, the AI designer for Civ 6 was ex-Starcraft, where he supposedly implemented BTs and brought that experience to civ 6. Big mistake.

Utility AI is the best system for TBS. Ironically again, Civ 5 scratched the surface and implemented a barebones Utility AI framework (they called it weight-based, and you can see it clearly in the code). Even if barebones, it is easy to see how superior vanilla civ 5 AI is to Civ6 BT AI. Gazebo and Ilteroi perfected the Utility AI system in Civ 5, and we got the infinitely superior AI in Vox.

There is no need for ML in civ yet. Utility AI is more than enough for the time being, and still provides powerful opponents while keeping turn times under control.

See here:

https://www.gamasutra.com/blogs/Jak...88/Are_Behavior_Trees_a_Thing_of_the_Past.php

Isn't Civ6 using a mix of utility AI and Behavior Trees ? (you can see in the exposed code that BT are very specific, somewhere between the strategical/tactical AI, with weighting still being used for "grand strategy")

Doesn't civ4/civ5/VP use BT at some levels ? (like tactical ?)

civ6 BT is clearly visible because it's exposed to modding, I don't remember how it was in civ4/civ5 DLL (well, I do remember that it was weight-based, what I don't remember is if it was only weight-based, and I've no looked at VP code but I'm curious)

That is an excellent point. I wonder if that would change TheMeInTeam's opinion of how well it would work for Civ? Hopefully, in the next episode, I'll remember to bring that up.

there is also the old but IMO still relevant Soren's video on fun vs good AI

and this thread too:
https://forums.civfanatics.com/threads/game-ai-net-based-machine-learning.627729/

To summarize:
- long path to reach a "winning condition"
- randomness (maps, events, combat results)
- incomplete information (fog of war, opponents unknown)
- complexity (terrain/feature types, combat rules, number of possible choices per turn, order of those choices, ...)
- evolving rules with each balance patches

 

[TheSpaceCowboy]

Why is PolyCast not available through Apple Podcasts?