Biological neural networks are interesting to study if you're designing artificial neural networks. They can provide insights to various solutions nature has found, which we can learn from. We can't program biological neural networks as they are driven by chemistry and physics rather than software, but various artificial neural network implementations have captured many aspects of them.

Another thing to note, artificial neural networks are great at certain problems, particularly problems where working out the solution details are extremely difficult, where data sets are far too large or too varied for humans to understand. Recognizing patterns in complex data is a common use. While that is a very useful field, it tends to be of limited use in games. The end results are useful for computers but cannot be easily modified, adjusted, or tuned.

Some elements may be useful like recognizing a gesture in a motion game, but for controlling elements inside the game they are a bad fit. You can easily memorize that a swipe-swipe-swipe motion means one thing, and a swipe-swipe-jab motion means something else, that's great. But deciding the ratio of workers, grunts, soldiers, warlocks, and dragons to build up, that's best left to humans designing the game using state machines, decision trees, and human-adjusted math functions. Designers can then adjust the level of difficulty to be difficult yet still winnable, a challenge which is fun and can be configured rather than brutal and optimal.

True machine learning AI's are miserable to play against. After they become well-tuned they take optimal paths, make optimal decisions, and can predict your every move. At least humans with aimbots and wall hacks still have human inefficiencies, an ideal AI is not fun to play against. When every action is optimized against you as the player, the AI makes no mistakes, takes actions the instant they are available, and makes the action perfectly, you've got to be just as perfect to be on par with them.

Moderation note: Removed some posts with off-topic idle banter and mild insults. Please keep posts on topic.

frob said:
Moderation note: Removed some posts with off-topic idle banter and mild insults. Please keep posts on topic.

In case my comment caused some response i have missed, let me clarify it was not meant insulting but just as a joke.

frob said:
Some elements may be useful like recognizing a gesture in a motion game, but for controlling elements inside the game they are a bad fit. You can easily memorize that a swipe-swipe-swipe motion means one thing, and a swipe-swipe-jab motion means something else, that's great. But deciding the ratio of workers, grunts, soldiers, warlocks, and dragons to build up, that's best left to humans designing the game using state machines, decision trees, and human-adjusted math functions.

Lets take the swipe example and turn it into a game mechanic of top down strategy game: The player gathers resources like some rocks, and he then can place them on terrain. If the placement forms a 'symbol of power', the player is rewarded with gaining power.
The larger the symbol is, the larger the reward.

Could this be an interesting mechanic, enabled by new technology of ML? Maybe, but there surely are much better examples we have not discovered yet. So i would not just say: ‘Nah - ML is good for upscaling, but not for the game itself.’
I think it's the other way around, and it's an opportunity we want to utilize. (just not me - i'm too old for this : )

… posting the same example i did in all of Calins AI threads again:

I doubt we can get this learning into the game on the client - it's probably doomed to remain offline on the training side, but…

• That's better anyways becasue we need predictable behaviour to design games
• This would allow to design games without a need to code gameplay code. Faster prototyping? Higher chance of inspiration and luck?
• We get NPCs that appear smarter, can interact with a more dynamic and interesting world.
• Offline training improvements could work on servers, maybe enabling games that evolve automatically over time. Manually tweaking balance might be enough.

I'm aware we want things under precise control, but just we are used to this being the only option at all so far, does not mean it is an requirement.

JoeJ said:
If the placement forms a 'symbol of power', the player is rewarded with gaining power. The larger the symbol is, the larger the reward. Could this be an interesting mechanic, enabled by new technology of ML? Maybe, but there surely are much better examples we have not discovered yet.

That's something that would be a poor fit for machine learning. There isn't much to reinforce or a way to get better, it is a binary state, either they made the shape or they didn't. During training a character could perform millions of pickup and placements and never create a single ‘symbol of power’.

Even if the system managed to generate one, the weights on an ANN machine don't have anything a game designer can tweak. They cannot turn a knob that says “do it bigger on higher difficulty”, or “do it more/less frequently”. The training phase runs for potentially many hours, and ultimately either stumbles up something good or doesn't, with nothing useful a human can do to change it.

Recognizing the shape would be a better fit for machine learning, and recognizer systems are commonplace in ML. You would need to classify a bunch of different shapes as being a proper ‘symbol of power’, and create a huge number of examples to learn from both demonstrating a ton of improper symbols and proper symbols, training until they are recognized. For that usage it is a simple matter of running the object positions through the ANN and getting a score back.

JoeJ said:
I doubt we can get this learning into the game on the client

A few games can and do implement ML algorithms on the client. They're just not typically used for game mechanics. As I mentioned in my earlier post, many games use them as gesture recognizers. With a little careful adjustment they can be generic to begin with, but then be designed to adapt to the person's personal gesture variations, becoming better able to recognize the person's unique motions.

But for game mechanics, where human designers need to fine tune values, humans adjust probabilities and frequencies and other weights, those tend to be a terrible fit for ML. Stick with state machines, behavior trees, and human adjustable math formulas.

The video you linked to even mentions the issues, albeit indirectly. “After millions of rounds of hide and seek, the finders begin to find a solution”, then “many million more rounds” and the seekers figure out they can use a ramp, and “many million rounds later” they learn to take away the ramp. The resulting ANNs are not easily adjusted values, but enormous arrays of training data that appears without rhyme or reason to a human game designer.