OpenTibia Weapon Sprites Generator (GAN)

[Otfan125]

okay sorry for spamming your thread, but I just remembered an old project I was doing and I want to hear your professional opinion on it:

Assuming this is a regular GAN network (log(D(X)) - log(1 - D(G(Z))), and the fact that we are dealing with relatively small data, what do you think of the idea of mixing some generated images with the real dataset based on some probability p after the generator is relatively doing well? So lets say after some time, the generator starts performing well (I'm thinking in terms of RL, but I hope you understand me) and generates 100 figures... then, we randomly pick 100*p images for 0<p<1 and mix it with the real data.... Whats the motivation? Well, some of the figures I've seen so far from this project of yours are relatively similar to the cipsoft ones (i mean... DUH!), and I think this method of adding some form of 'noise samples, if you will' to the real data can maaaaaaaybe have the generator produce more eccentric images? Sorry, even I lost track of my thinking during typing, anyways, i hope this is interesting to you as it is to me

Post automatically merged: Aug 2, 2020

I did forget to mention that the generated images that were picked randomly via probability p should get replaced as the generator begins producing "better" images...

 

[Night Wolf]

As for your other message @Otfan125, I'm afraid it doesn't work that way, I will edit this comment later with a more detailed explanation but right now I'm in the cellphone

Post automatically merged: Aug 2, 2020

okay sorry for spamming your thread, but I just remembered an old project I was doing and I want to hear your professional opinion on it:

Assuming this is a regular GAN network (log(D(X)) - log(1 - D(G(Z))), and the fact that we are dealing with relatively small data, what do you think of the idea of mixing some generated images with the real dataset based on some probability p after the generator is relatively doing well? So lets say after some time, the generator starts performing well (I'm thinking in terms of RL, but I hope you understand me) and generates 100 figures... then, we randomly pick 100*p images for 0<p<1 and mix it with the real data.... Whats the motivation? Well, some of the figures I've seen so far from this project of yours are relatively similar to the cipsoft ones (i mean... DUH!), and I think this method of adding some form of 'noise samples, if you will' to the real data can maaaaaaaybe have the generator produce more eccentric images? Sorry, even I lost track of my thinking during typing, anyways, i hope this is interesting to you as it is to me

Post automatically merged: Aug 2, 2020

I did forget to mention that the generated images that were picked randomly via probability p should get replaced as the generator begins producing "better" images...

so basically the main problem that we have to do this approach is that the final result still contains a lot of artefacts and noise, so doing this would insert more and more noise in the model.

Also, the results we can find by exploring the latent space are far more eccentric (in fact you can choose how eccentric you want them) than just outputing random values that could pottentialy be more similar to existing ones in our base and produce problems such as model collapsing (when we train with samples that are very alike and the model starts producing only copies of trained samples.

 

[Otfan125]

Hey Night, I have a question for you... do you think its possible to have an AI player in an OT that hunts, sells, etc.? (basically like a human would)
This is interesting to me since I'm in the sub field of Reinforcement learning (or at least, I used to be, got hired by molecular computation nerds that give me too much work!)...... I think it would be interesting to see how different RL agents interact with each other in the game. I understand it is going to be quite the challenge, but I bet its possible... Maybe have different networks for different tasks (hunting, gathering, etc.) and the agent can switch between those tasks/networks... I bet that having, say, 100 agents in an OT would not take too much gpu power to run...

 

[Night Wolf]

Hey Night, I have a question for you... do you think its possible to have an AI player in an OT that hunts, sells, etc.? (basically like a human would)
This is interesting to me since I'm in the sub field of Reinforcement learning (or at least, I used to be, got hired by molecular computation nerds that give me too much work!)...... I think it would be interesting to see how different RL agents interact with each other in the game. I understand it is going to be quite the challenge, but I bet its possible... Maybe have different networks for different tasks (hunting, gathering, etc.) and the agent can switch between those tasks/networks... I bet that having, say, 100 agents in an OT would not take too much gpu power to run...

we are still bounded by gpu power, training a human-like behavior on tibia would require a lot of time because there's too many things to learn.
Reinforcement learning wouldn't be the best approach as it would require us to determine what is the reinforcement factor (time alive perhaps?)
But that can make the bot stop standing in an area until it's killed by another player, that would be a valid strategy...

 

[Otfan125]

I like your answer, and as a researcher, I have many directions and things I want to say on the topic (ideas, theoretical ones, etc.) though, I think we'd require a whiteboard for these questions and, well, I don't think it would be time productive to attempt to describe them here... I think its best if maybe I prepare a document formally describing these ideas once I get the chance... thanks

 

[Night Wolf]

feel free to hit me on discord if you want, Night Wolf#8602 we can discuss the details there and bring the highlights here