Reliable automatic evaluation of dialogue systems under an interactive environment has long been overdue. An ideal environment for evaluating dialog systems, also known as the Turing test, needs to involve human interaction, which is usually not affordable for large-scale experiments. Though researchers have attempted to use metrics (e.g., perplexity, BLEU) in language generation tasks or some model-based reinforcement learning methods (e.g., self-play evaluation) for automatic evaluation, these methods only show a very weak correlation with the actual human evaluation in practice. To bridge such a gap, we propose a new framework named ENIGMA for estimating human evaluation scores based on recent advances of off-policy evaluation in reinforcement learning. ENIGMA only requires a handful of pre-collected experience data, and therefore does not involve human interaction with the target policy during the evaluation, making automatic evaluations feasible. More importantly, ENIGMA is model-free and agnostic to the behavior policies for collecting the experience data (see details in Section 2), which significantly alleviates the technical difficulties of modeling complex dialogue environments and human behaviors. Our experiments show that ENIGMA significantly outperforms existing methods in terms of correlation with human evaluation scores.

A model-based reinforcement learning (MBRL) agent gradually learns a model of the environment as it interacts with it, and uses the learned model to plan and find a good policy. This can be done by planning with samples coming from the model, instead of or in addition to the samples from the environment, e.g., Sutton (1990); Peng & Williams (1993); Sutton et al. (2008); Deisenroth et al. (2015); Talvitie (2017); Ha & Schmidhuber (2018). If learning a model is easier than learning the policy or value function in a model-free manner, MBRL will lead to a reduction in the number of required interactions with the real-world and will improve the sample complexity of the agent. However, this is contingent on the ability of the agent to learn an accurate model of the real environment. Therefore, the problem of learning a good model of the environment is of paramount importance in the success of MBRL. This paper addresses the question of how we can approach the problem of learning a model of the environment, and proposes a method called policy-aware model learning (PAML). The conventional approach to model learning in MBRL is to learn a model that is a good predictor of the environment. If the learned model is accurate enough, this leads to a value function or a policy that is close to the optimal one. Learning a good predictive model can be achieved by minimizing some form of a probabilistic loss.