Show whole topic
Jan 02, 2016 11:55 pm
Jan 02, 2016 11:55 pm
Subject: Re: Ai implementation
Hello everybody and happy new year!
I was playing MAX over the holiday and I remembered the thread. I think the major issue is not the architecture (whether we do in-process or out-process, using LUA or C++, etc ) but the major issue is how to do it with regard of algorithmic implementation. The game is very complex, as it requires an interconnected "threads" of economy, exploration and military strategy. They are interconnected because focusing on economy only will make the player vulnerable to early attacks, while focusing on military expenditure will affect the player on the long run, if unable to defend quickly the adversaries.
What I was wondering these days is if you have any large number of games stored (as a set of events) and in case you do, this may be fed to a neural network. I can contribute the NN and the initial training. I don't know how well can be trained such a player, but I think it is worth trying. The game is highly stateful (one cannot build if no materials are available) or it requires a long chain of events strategy therefore I don't know if the NN will be able to follow up such a chain or rather jump from one strategy to another.
Also, the neural network will "try" at least in the initial phase to build or move items to invalid places, the server needs to be able to reject such requests.
On the other hand, being able to "cheat" with regard of exploration, adversary position and assets would greatly boost the AI.
Cheers,
Razvan
Hello everybody and happy new year!
I was playing MAX over the holiday and I remembered the thread. I think the major issue is not the architecture (whether we do in-process or out-process, using LUA or C++, etc ) but the major issue is how to do it with regard of algorithmic implementation. The game is very complex, as it requires an interconnected "threads" of economy, exploration and military strategy. They are interconnected because focusing on economy only will make the player vulnerable to early attacks, while focusing on military expenditure will affect the player on the long run, if unable to defend quickly the adversaries.
What I was wondering these days is if you have any large number of games stored (as a set of events) and in case you do, this may be fed to a neural network. I can contribute the NN and the initial training. I don't know how well can be trained such a player, but I think it is worth trying. The game is highly stateful (one cannot build if no materials are available) or it requires a long chain of events strategy therefore I don't know if the NN will be able to follow up such a chain or rather jump from one strategy to another.
Also, the neural network will "try" at least in the initial phase to build or move items to invalid places, the server needs to be able to reject such requests.
On the other hand, being able to "cheat" with regard of exploration, adversary position and assets would greatly boost the AI.
Cheers,
Razvan