The main goal of the adaptive abstractor is to determine a suitable state abstraction based on the learnt communication protocol, which reduces the size of the state space to be explored (the semantic problem), which in turn helps the agents achieve their goal, without much degradation in the policy performance (the effectiveness problem). Unlike [1], the proposed approach does not assume any prior knowledge of an expert policy to learn the abstraction. Moreover, the size of the abstracted space is not determined a priori.
Interesting. Use the skills learned to determine 'what is necessary' in the state representations? It is basically like the strategy space