Relative Variational Intrinsic Control

In the absence of external rewards, agents can still learn useful behaviors by identifying and mastering a set diverse skills within their environment. Existing skill learning methods use mutual information objectives to incentivize each be distinguishable from rest. However, if care is not taken constrain ways in which are diverse, trivially sets arise. To ensure diversity, we propose novel objective, Relative Variational Intrinsic Control (RVIC), incentivizes that how they change agent's relationship its The resulting tiles space affordances available agent. We qualitatively analyze on multiple environments show RVIC more than discovered existing hierarchical reinforcement learning.