This paper presents 'SimpleDS', a simple and publicly available dialoguesystem trained with deep reinforcement learning. In contrast to previousreinforcement learning dialogue systems, this system avoids manual featureengineering by performing action selection directly from raw text of the lastsystem and (noisy) user responses. Our initial results, in the restaurantdomain, show that it is indeed possible to induce reasonable dialogue behaviourwith an approach that aims for high levels of automation in dialogue controlfor intelligent interactive agents.
展开▼