We propose a DQN based technique to generalize the concept of skills. In previous papers, we have seen attempts to hand-craft skills (Chen Tessler, 2016). The idea of skills is that skills reduce the effective size of the MDP by enabling the agent to plan at a higher level. More than that, every specific task will be performed in the best way. However, skills are specific for every domain. In this paper, we attempt to create an agent that will have the advantage of using skills that are both meaningful and general enough to be applied to every environment.
We created the agent based on Dopamine framework and trained it on Atari games. The skills were implemented as sub-agentswithdifferentproperties.Indoingsowewantedto achieve a wide range of behaviors, every agent will be useful inspecificstates,andontopofthat,theskillsareextractedautonomously. Unlike classic DQN in which the agent chooses an action in every step, our agent will choose a different subagent every fixed number of steps.