StarCraft® II

Genetic Programming

Posts: 1,064
What could blizzard do with genetic programming? Could they have AI that improves its strategy after every game? Or could they implement ways to use this concept in the map editor to provide map makers with more powerful tools?
Reply Quote
Posts: 30
Genetic programming has proven very unsuccessful. The truth is random changes take a lot of time beofre anything good is selected. There are a lot of other better AI techniques for this. If i had to pick a famous one i would go with minmax.
Reply Quote
Posts: 1
In theory it would be a very interesting way to tune the vs ai games. You could have all the current settings of difficulty, and a new one that remembers how you play, and changes, etc. However, programing an actual good one that works and learns well, would be a very difficult challenge.
Reply Quote
Posts: 68
one step away from true A.I
Reply Quote
Posts: 6,061
The AI would have to be able to learn from all game data. It would have to somehow analyze replays and correlate build times, costs, damage output which units died the most, which were harder to protect, hp, shields, costs of upgrades, and it would then need a way to take all that and determine the most efficient builds or the best ways to counter. It would have to be a terrific scout always using overlords, changlings and even an army unit or two just to see for a moment what it's up against.

The AI should be centralized though and not just relative to each player though. Maybe both ya sure, but this is data it would need to collect thousands of times from every single melee game played on Every replay. And it would need a way to determine the faults of other players such as inefficiency with resource management and use that information to interpret how certain encounters don't mean exactly what they mean or that they do exactly mean what they mean in relation to how resources were spent to reach those points in the armies prior to battle.

Such an AI would basically need to be programmed to have complete "sensory" input and teach itself and then adjust it's own code to adapt and evolve. Whoever creates this kind of AI could doom us all ... and I don't mean just the game world if a non game AI was created.
Reply Quote
Posts: 6,061
Has anyone even to date created a self sustaining AI?
Reply Quote
Posts: 560
05/03/2011 05:06 PMPosted by Lykos
Has anyone even to date created a self sustaining AI?


Its skynet on PCP
Reply Quote
Posts: 425
Using genetic algorithms would be too complex and suffer from a strong inductive bias. There is no way to use automated learning algorithms to create a self learning AI. The closest you can get is to feed the AI with lots of replays and have it use bayesian learning using the build orders as input (and making it map specific).

Automated learning would never work on starcraft because this game is heavily positional and based around micro.

However, if you want to see how to make AI godlike, check this videos:

Reply Quote
Posts: 3
Genetic algorithms have been successfully applied to subsets of the game.

Build Orders

Targeting (SC 1)

If you can define the fitness function and twiddly bits well enough genetic algorithms have a lot of promise. I think genetic algorithms have a bright future in continuing to address subsets of game AI algorithms.
Reply Quote
Posts: 178
True, they might make another A.I. that a human player would have no hope of winning against because it would have always perfect macro and godly micro. The only disadvantage to an A.I. is that it won't be able to deal with other people's builds that are weird as well. Yeah you can programm it to counter opposing compoistions but things like a cannonr suh would start to throw it off and it would constantly get to a certain point. Like an A.I. with perfect micro/macro can win the game with purely blink stalkers since no stalker would ever die it would just keep blinking back as shields are depleted.
Reply Quote
Posts: 751
Genetic engineering for spanshwa as an A.I
Reply Quote
Posts: 186
The thing to remember about video game AI is it's meant to be fun, not to be optimal. Think about a bot in a deathmatch style game. It would be easy to create a bot that had perfect vision and perfect aim, but that game would be frustrating to the player since the player couldn't win. Likewise, a Starcraft AI could be designed to perfectly counter an opponents build, but the player would hate playing against it since it would be unbeatable.

<offtopic> Perhaps that would be a good way to measure if a build is OP. If a build could beat an unbeatable AI then the build is by definition OP</offtopic>
Edited by TurdFurgison on 5/4/2011 6:12 AM PDT
Reply Quote
Posts: 1,014
Maybe not generic programming but other AI methods would be fine. For example, it would be a Ai system that process every game in and learn about streategies, so they could improve in time and take care about metagame shifts.
Reply Quote
Posts: 42
why would we need genetic programing to make an ai? our brains isnt genetic programing right? just regular neurons acting like transistors
Reply Quote

Please report any Code of Conduct violations, including:

Threats of violence. We take these seriously and will alert the proper authorities.

Posts containing personal information about other players. This includes physical addresses, e-mail addresses, phone numbers, and inappropriate photos and/or videos.

Harassing or discriminatory language. This will not be tolerated.

Forums Code of Conduct

Report Post # written by

Explain (256 characters max)
Submit Cancel