Efficient Exploration for Dialog Policy Learning with Deep BBQ Networks \& Replay Buffer Spiking.

When rewards are sparse and action spaces large, Q-learning with $epsilon$-greedy exploration can be inefficient. This poses problems for otherwise promising applications such as task-oriented dialogue systems, where the primary reward signal, indicating successful completion of a task, requires a complex sequence of appropriate actions. Under these circumstances, a randomly exploring agent might never stumble upon a successful outcome in reasonable time. We present two techniques that significantly improve the efficiency of exploration for deep Q-learning agents in dialogue systems. First, we introduce an exploration technique based on Thompson sampling, drawing Monte Carlo samples from a Bayes-by-backprop neural network, demonstrating marked improvement over common approaches such as $epsilon$-greedy and Boltzmann exploration. Second, we show that spiking the replay buffer with experiences from a small number of successful episodes, as are easy to harvest for dialogue tasks, can make Q-learning feasible when it might otherwise fail.