UCB and InfoGain Exploration via $\boldsymbol{Q}$-Ensembles

We show how an ensemble of $Q^*$-functions can be leveraged for more effective exploration in deep reinforcement learning. We build on well established algorithms from the bandit setting, and adapt them to the $Q$-learning setting. First we propose an exploration strategy based on upper-confidence bounds (UCB). Next, we define an "InfoGain" exploration bonus, which depends on the disagreement of the $Q$-ensemble. Our experiments show significant gains on the Atari benchmark.