W:A Neural Network

[Ytrojan]

Someone should make a Worms Armageddon Neural Network, either to solve missions or to play schemes during local multiplayer. Here are some advantages over the regular CPU
-Being able to play more than just battle schemes
-Less idiocy
-Being able to play like a regular Human player
-BRSolver for the public
Think of it as a sort of BRSolver, except for all schemes.

[Xrayez]

Many W:A's custom schemes have human defined rules and the scheme format doesn't allow to create winning conditions, besides killing all worms. In order to play the scheme, one generally has to learn human readable rules, which is virtually impossible for the neural network to figure out. The only feasible schemes are schemes like Intermediate/Normal. The neural network could analyze the vast amount of Intermediate replays and come up with a general intelligence. 

[zippeurfou]

Not really X-ray. You can put objective functions in RL and sequence them. I.e get crate and then attack.
You also can’t use the replay as you need to be able to capture the input

[Xrayez]

Not really X-ray. You can put objective functions in RL and sequence them. I.e get crate and then attack.

I think that would lead to implicitly defining rules for the agents manually. What I'm talking about is that it would be really good if W:A had mechanisms to define rules (along with environment) within a scheme so that RL methods could be conveniently applied without having to worry about creating complex and chained objective functions (if I understand you correctly).

You also can’t use the replay as you need to be able to capture the input

Aren't W:A replays contain the actual input to simulate the game? Both game states and inputs could be retrieved.

[zippeurfou]

Not really X-ray. You can put objective functions in RL and sequence them. I.e get crate and then attack.

I think that would lead to implicitly defining rules for the agents manually. What I'm talking about is that it would be really good if W:A had mechanisms to define rules (along with environment) within a scheme so that RL methods could be conveniently applied without having to worry about creating complex and chained objective functions (if I understand you correctly).

You also can’t use the replay as you need to be able to capture the input

Aren't W:A replays contain the actual input to simulate the game? Both game states and inputs could be retrieved.

I agree with your first part but sadly most RL work like that.
For the second part it seems like you're right (http://worms2d.info/Replay_file) still far from trivial to do.