H-DRLN

This is in development stage

The H-DRLN (Hierarchical Deep RL Network) is a novel architecture to incorporate options (skills) within the DQN. For further explanation see the website.

Graphic view of created skill clusters

We use the work done by Zahavy et al and adapt the interface they created in order to graphically view the clusters on the tSNE plot.

References

• A Deep Hierarchical Approach to Lifelong Learning in Minecraft - Tessler et al
• Human-level control through deep reinforcement learning. Nature 518(7540):529–533 Mnih et al
• Graying the black box: Understanding dqns. Proceedings of the 33th international conference on machine learning (ICML) - Zahavy et al