Deriving Subgoals Using Network Distillation

Student thesis: Master thesis (including HD thesis)

  • Nikolaj Ljørring
  • Lars Svane Jensen
  • Aryan Mohammadi Landi
4. term, Computer Science, Master (Master Programme)
Sparsely rewarded environments can be challenging
for deep reinforcement learning to understand and even harder
to master. Hierarchical reinforcement learning shows promising
ways of constructing subgoals, that are more understandable
to the agent. Subgoal construction is a slow process to do
autonomously, we therefore propose a new method of finding
and constructing subgoals. We present a more time-efficient
comparison method for subgoal creation. We propose a novel
distributed training framework to increase the throughput of
the agent. The framework indicates increased data gathering but
decreased learning compared to a non-distributed counterpart.
Publication date11 Jun 2021
Number of pages12
ID: 414297993