Sample-Efficient Reinforcement Learning with Maximum Entropy Mellowmax Episodic Control
This work addresses sample efficiency for reinforcement learning practitioners, but it is incremental as it builds on existing episodic control methods with a novel exploration strategy.
The paper tackled the problem of sample inefficiency in reinforcement learning by proposing maximum entropy mellowmax episodic control (MEMEC), which uses a Boltzmann policy with state-dependent temperature for exploration. The result showed that MEMEC outperformed other methods on classic RL environments and Atari games, achieving faster learning and higher final rewards.
Deep networks have enabled reinforcement learning to scale to more complex and challenging domains, but these methods typically require large quantities of training data. An alternative is to use sample-efficient episodic control methods: neuro-inspired algorithms which use non-/semi-parametric models that predict values based on storing and retrieving previously experienced transitions. One way to further improve the sample efficiency of these approaches is to use more principled exploration strategies. In this work, we therefore propose maximum entropy mellowmax episodic control (MEMEC), which samples actions according to a Boltzmann policy with a state-dependent temperature. We demonstrate that MEMEC outperforms other uncertainty- and softmax-based exploration methods on classic reinforcement learning environments and Atari games, achieving both more rapid learning and higher final rewards.