Such an agent is typically characterised by: (i) a finite set of states 6 S {s i} that describe all possible situations in the environment; (ii) a finite set of actions A {a j} to change in the environment from one situation to another; (iii) a state transition function T (s,a,s null) that specifies the next state s null for having taken action a in the current state s; (iv) a reward function R (s,a,s null) that specifies a numerical value given to the agent for taking action a in state s and transitioning to state s null; and (v) a policy π: S A that defines a mapping from states to actions [2, 30]. The goal of a reinforcement learning agent is to find an optimal policy by maximising its cumulative discounted reward defined as Q (s,a) max π E[r t γr t 1 γ 2 r t 1 ... s t s,a t a,π ], where function Q represents the maximum sum of rewards r t discounted by factor γ at each time step. While a reinforcement learning agent takes actions with probability Pr ( a s) during training, it selects the best action at test time according to π (s) arg max a A Q (s,a). A deep reinforcement learning agent approximates Q using a multi-layer neural network [31]. The Q function is parameterised as Q(s,a; θ), where θ are the parameters or weights of the neural network (recurrent neural network in our case). Estimating these weights requires a dataset of learning experiences D {e 1,...e N} (also referred to as'experience replay memory'), where every experience is described as a tuple e t ( s t,a t,r t,s t 1). Inducing a Q function consists in applying Q-learning updates over minibatches of experience MB {( s,a,r,s null) U (D)} drawn uniformly at random from the full dataset D . This process is implemented in learning algorithms using Deep Q-Networks (DQN) such as those described in [31, 32, 33], and the following section describes a DQN-based algorithm for human-chatbot interaction.

Training chatbots using the reinforcement learning paradigm is challenging due to high-dimensional states, infinite action spaces and the difficulty in specifying the reward function. We address such problems using clustered actions instead of infinite actions, and a simple but promising reward function based on human-likeness scores derived from human-human dialogue data. We train Deep Reinforcement Learning (DRL) agents using chitchat data in raw text---without any manual annotations. Experimental results using different splits of training data report the following. First, that our agents learn reasonable policies in the environments they get familiarised with, but their performance drops substantially when they are exposed to a test set of unseen dialogues. Second, that the choice of sentence embedding size between 100 and 300 dimensions is not significantly different on test data. Third, that our proposed human-likeness rewards are reasonable for training chatbots as long as they use lengthy dialogue histories of >=10 sentences.

The amount of dialogue history to include in a conversational agent is often underestimated and/or set in an empirical and thus possibly naive way. This suggests that principled investigations into optimal context windows are urgently needed given that the amount of dialogue history and corresponding representations can play an important role in the overall performance of a conversational system. This paper studies the amount of history required by conversational agents for reliably predicting dialogue rewards. The task of dialogue reward prediction is chosen for investigating the effects of varying amounts of dialogue history and their impact on system performance. Experimental results using a dataset of 18K human-human dialogues report that lengthy dialogue histories of at least 10 sentences are preferred (25 sentences being the best in our experiments) over short ones, and that lengthy histories are useful for training dialogue reward predictors with strong positive correlations between target dialogue rewards and predicted ones.

To ensure satisfactory user experience, dialog systems must be able to determine whether an input sentence is in-domain (ID) or out-of-domain (OOD). We assume that only ID sentences are available as training data because collecting enough OOD sentences in an unbiased way is a laborious and time-consuming job. This paper proposes a novel neural sentence embedding method that represents sentences in a low-dimensional continuous vector space that emphasizes aspects that distinguish ID cases from OOD cases. We first used a large set of unlabeled text to pre-train word representations that are used to initialize neural sentence embedding. Then we used domain-category analysis as an auxiliary task to train neural sentence embedding for OOD sentence detection. After the sentence representations were learned, we used them to train an autoencoder aimed at OOD sentence detection. We evaluated our method by experimentally comparing it to the state-of-the-art methods in an eight-domain dialog system; our proposed method achieved the highest accuracy in all tests.