I'm running A3C (Asynchronous Actor-Critic Agents) to learn a game where an agent needs to catch 3 rewards. The input of my network, among other things, is the relative position of the 3 rewards against the agent.
However, as the agent catches these rewards, my network can't just decrease its input size. What are ways of handling this?
My current solutions have been:
set remaining features to zero - have tried, bad results
just show the closest reward - loss of information, might not be optimal
repeat existing reward positions to remaining features - we are giving more emphasis to some rewards (the ones that are repeated) over others