I'm working on a network restoration problem using reinforcement learning. Each step involves restoring a limited number of sensors. To represent the action space, I initially employed a combinatorial approach, creating a discrete action space where each action corresponds to a combination of sensors to be restored. For instance, if I have 21 sensors to restore and can restore 4 sensors per step, I generate an action space with 21 choose 4 (21C4) combinations. While this approach works well for small-scale systems, I'm encountering challenges when dealing with larger networks.
I have attempted to solve this restoration problem by applying Deep Q-Networks (DQN), and the model performs adequately for smaller systems. However, as the scale of the network increases, the computational demands become prohibitive.
I'm seeking advice on how to efficiently formulate the action space for this large-scale network restoration problem. Are there alternative methods or strategies that can be employed to create a more scalable action space representation? or If there any other way I can formulate the problem? Thank you!