Flexible subgoal representations towards object-orientedness
The subgoal representations should be object-oriented and at the same time flexible wrt. the size and sub-vector of the observation space. For example, in the Ant Environment, it is important to have velocities in the subgoals, but not in the end-goals. Therefore, we need two features:
- set size of subgoal vector for each layer as a hyperparameter option
- set obs2goal mapping either manually for each layer (as required for the velocities in the ant env), or let the actor of the higher layer determine the mapping. In both cases, we can test two versions:
- Set the subgoal vector freely, i.e., it could be the x-axis of the location of an object and the y-axis of the velocity.
- Have a fixed grouping size of 3, so that the mapping determines either e.g. location or velocity or rotation of an object. However, this requires to make the observation vector divisible by 3 in a meaningful way.
Edited by Manfred Eppe