Originally Posted by thEsp
There's something I did not understood. What neural networks can be used in pawn for? As Crazy said, can NPCs (or bots) be programmed using the same concept?
When you have a problem that maybe you don't know how solve it then you can just send some labeled data
to the neural network, so after the learning process you will can achieve the problem. Obviously, you also can do it in a imperative way with if, else, etc but sometimes there is better and simple ways
I have your suggestion in mind from the beginning but this contribution is still in development
Originally Posted by Bugsy
Agree, this is basically useless to the AMX-X community, IMO. And I don't mean this in a bad/negative way, but nobody is going to know how to apply this to a game-play plugin. You should create a plugin that uses this for actual game play purposes, not just a plugin that does calculations and prints them.
. Especially at the beginning, it was not easy even to make a simple plugin. Little by little I'm giving it improvements to be easy to use.
For example, the last optimizations are important to "think" in real time, for example, it's still useful for NPC. Nobody wants a server that consumes many resources. In addition, they will want to have many NPCs at the same time, each optimization is important.
Also, the last update greatly improved the flexibility of use and keeping good performance
Originally Posted by EFFx
Tbh, the main purpose for A.I is letting it make basically or literally what normal humans/players can do and even better/faster. As Bugsy said, it could be kinda useless to programmers even Veteran ones who didn't understand everything at all. So I'm another person who says it: That's freaking cool, but could be useless since we cannot apply this on good purposes as NPC's or whatever we want to make but instead of registering pre-words or actions that can't be done, it'll learn and avoid in the next time.
That's my vision of A.I, maybe my vision is too hard for AMXX since it's not that supportable, I guess.
What I marked in bold, I was thinking about it and that's why I detailed all the code (.inc file).
In addition, I always improve the API to reduce complexity and this breaks compatibility
so I don't bother doing an advanced plugin now, I could but it's too much maintenance and I have the same post in Spanish (amxmodx-es) and English.
I think that AMXX is perfectly compatible, there are many mathematical details that I don't know
and that could be useful for optimization in the learning process, for example, adaptable learning rate
. My knowledge is purely self-taught since always so I need time.
Also, I never do developments in SourcePawn (was my first language but I'm with others things since a lot of time) so I'm not so capable of doing NPC now, I plan to use something already developed as base: https://amxmodx-es.com/Thread-NPC-Cucarachas
but takes some time because I have more priorities.
Originally Posted by thEsp
I don't agree with the concept of machines being able to learn. Something that does not have senses,feelings and ability to predict cannot evolve.
Yeah, even veterans may not be aware of this deep section. I've thought some ideas about usage of A.I (in Pawn) such as an verify system (similar to captcha). Turns out there can be achieved something via this concept, BUT not with resources we maintain right now. You can't do much with some decimal numbers translated to regular integers. At least these are my thoughts.
I disagree a bit, there are many types of neural networks. For example, there is reinforced learning
, something like dopamine, when you do something well it's reinforced by the positive result and vice versa when the result was not the expected. In a game when the bot win or lose then positive or negative reinforcement is applied
Anyway, don't expect them to self-evolve from nothing, everything has to be pretty specifically indicated but you'd be surprised :p
By the way, this library is not that kind of neural network
. It's deep feed forward
You said "You can't do much with some decimal numbers translated to regular integers" but, in fact, yes. It's how any neural network is handled.
Maybe you will not apply neural networks in AMXX but yes in other languages, all this is just a bit of knowledge that can be used in more places.
TensorFlow developed by Google:
That is an extremely basic example (typical xor) but they have the most completed and great library to do anything with a lot of documentation, examples, tutorials, videos, etcetera.