Currently, I’m designing a neural network that works with reinforcement learning. In summary, the agent takes in information about itself and nearby agents and, in conjunction with global world information, makes a decision.
I’m currently thinking of implementing this as a LSTM to take in information about itself and a variable number of nearby agents and a feedforward neural network to combine the information from the LSTM output and global world information to produce an action.
Would this approach be sufficient to produce meaningful results? I thought that another approach would be to take in the global world information and each agent at each LSTM cell, though it may use much more resources (resources during forward propagation are a main concern with this project). Also, if the second approach is used, how would I be able to link the inputs to outputs if they had different shapes (attempting to learn without a library)? How would I be able to map an input with shape (1, x, 6) to (1, 1, 4) or (1, 4).