Collaborative Multi-Agent Dialogue Model Training Via Reinforcement\n Learning

We present the first complete attempt at concurrently training conversational\nagents that communicate only via self-generated language. Using DSTC2 as seed\ndata, we trained natural language understanding (NLU) and generation (NLG)\nnetworks for each agent and let the agents interact online. We model the\ninteraction as a stochastic collaborative game where each agent (player) has a\nrole ("assistant", "tourist", "eater", etc.) and their own objectives, and can\nonly interact via natural language they generate. Each agent, therefore, needs\nto learn to operate optimally in an environment with multiple sources of\nuncertainty (its own NLU and NLG, the other agent's NLU, Policy, and NLG). In\nour evaluation, we show that the stochastic-game agents outperform deep\nlearning based supervised baselines.\n