The AI experiment, creating Algorythms
-
I was kinda curious about self learning algorithms. I asked one if they are able to learn Syntax yet. It said it could as long as it had a manual.
The one I tried was GPT3 but it isn't open enough to learn from the get go.
Anyway this is a crazy idea but shouldn't be too hard to achieve...Then Secondly I came to the idea of learning Algorithms and wondered if it's possible to write one that's just concept-ish close in FUZE.
Anyone having experience in that direction?
-
I started to look into ML to find out if I can use it for a single player mode of Stroll Wars. The applied algorithm in the example below is a Q-Learning algorithm. It is not a deep learning version yet and I am not sure if I did the implementation right. Glad to learn from a discussion here.
Share ID: HNZR3MNDN8
I ported the Python implementation from Soren D of "catch the cheese".
All possible input variables are stored in an array, and an action to change the state is randomly chosen. If the action leads to a success (like catching the cheese), the action for this input variable states is rewarded. After some time, the array contains the most efficient (by 9x%) solution. But a certain random component has to be applied at any time, to avoid that the algorithm gets stuck at a path that leads to a solution, but is not the best yet. -
@Mechanical It's possible but getting a training set fed into Fuze sounds very impractical right now :/
-
I wondered if it's possible to convert input into output like a chat bot.
I know there are some functions to give input like text while a program runs.
Maybe someone can explain how to achieve that an algorithm learns what to say. Or to find contradictions like GPT3.
I'm amazed of the new AI tech especially GPT3. It's intelligence skyrocketed over the last months and soon it will be in every OS as an assistant. Microsoft bought the implementation for it. Of course that's a huge upgrade to cortana.
The days will come, where we are coding and constructing by just defining what we want to have. AI will assist us in these things.
-
@Mechanical obviously your first drawback is not having the whole web of data being fed to the algorithm to learn from but this posses a fun challenge involving phycology.
List of possible input data for the algorithm:
How long does it take the user to click through a set of images, sounds.
Then you could check reaction times with simple tests,
then store all that data and it will know what background image to avoid based on what image you clicked off the fastest, and it will know your reation time.
Could be handy having some thing like this before a game starts or whatever.
Sorry if I missed the point of your post ... That's my take on it. -
It seems like this thread is sort of overthinking things: you could do "machine learning" by just writing the player's actions and the outcomes to a file, then have your AI logic read that file, and use branching "if" statements. Overall, the AI logic would be like, "the player tends to go left more than right, so err that way", or "half the time the player went up the middle they scored, so defend there."
You could technically use this approach on much weaker hardware than the Switch. I've even thought of making "adaptive" AI in an Atari ST game or something, I think it'd work fine.
