How would one select moves in a complex game that has multiple screens. For instance, how would you handle things like item screens and skill point selection screens when implementing a reinforcement learning agent to play a game. What kinds of output spaces will be required of the network? For instance, a simple dungeon crawler can deal with controls by outputting the movement and weapon/skill controls. However, what about dealing with menus like equipment layouts etc. How do we modify the network to deal with this as well?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
An MDP (markov decision process) or POMDP (partially observable MDP) are two popular mathematical abstractions which are used to describe how an agent interacts with the environment. Almost all games or tasks can be formulated as either an MDP or a POMDP.
Therefore, when people develop RL algorithms, they don't design them for a specific game or environment. Instead, RL algorithms are usually designed for a general class of problems (say, MDPs with unknown dynamics).
Multiple screens or interfaces, and the controls to move between them doesn't pose any new challenge, because they can be pretty easily added onto the environment in the (PO)MDP framework. So fundamentally, there is nothing that needs to be changed.
