# A Deep Reinforcement Learning Chatbot (Short Version)

**Authors:** Iulian V. Serban, Chinnadhurai Sankar, Mathieu Germain, Saizheng Zhang, Zhouhan Lin, Sandeep Subramanian, Taesup Kim, Michael Pieper, Sarath Chandar, Nan Rosemary Ke, Sai Rajeswar, Alexandre de Brebisson, Jose M. R. Sotelo, Dendi Suhubdy, Vincent Michalski, Alexandre Nguyen, Joelle Pineau, Yoshua Bengio  
**arXiv:** [1801.06700](https://arxiv.org/abs/1801.06700)  
**Date:** 2018-01-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.06700v1/

## Key points

- The Montreal Institute for Learning Algorithms (MILA) has developed MILABOT, a deep reinforcement learning chatbot for the Amazon Alexa Prize competition.
- MILABOT engages in small talk conversations with humans using both speech and text.
- The system utilizes natural language generation and retrieval models, including neural network and template-based models.
- MILABOT has been trained to select appropriate responses from its ensemble of models using reinforcement learning techniques on crowdsourced data and real-world user interactions.
- A/B testing with real-world users has shown significant improvements in MILABOT's performance compared to other systems.
- This research highlights the potential of integrating ensemble systems with deep reinforcement learning as a promising approach for developing practical open-domain conversational agents.
- The paper presenting this work provides additional details about the development process and evaluation results at the NIPS 2017 Conversational AI workshop.
- It demonstrates how MILABOT effectively uses reinforcement learning techniques to improve its performance in engaging in small talk conversations with humans using both speech and text.
- Combining ensemble systems with deep reinforcement learning can be an effective approach for creating practical open-domain conversational agents.

## Summary

The Montreal Institute for Learning Algorithms (MILA) has developed MILABOT, a deep reinforcement learning chatbot, for the Amazon Alexa Prize competition. MILABOT is designed to engage in small talk conversations with humans using both speech and text. The system utilizes a combination of natural language generation and retrieval models, including neural network and template-based models. By leveraging reinforcement learning techniques on crowdsourced data and real-world user interactions, MILABOT has been trained to select appropriate responses from its ensemble of models. The system's performance has been evaluated through A/B testing with real-world users, demonstrating significant improvements compared to other systems. This research highlights the potential of integrating ensemble systems with deep reinforcement learning as a promising approach for developing practical open-domain conversational agents. The paper presenting this work provides additional details about the development process and evaluation results at the NIPS 2017 Conversational AI workshop. It demonstrates how MILABOT was able to effectively use reinforcement learning techniques to improve its performance in engaging in small talk conversations with humans using both speech and text. Moreover, it shows that combining ensemble systems with deep reinforcement learning can be an effective approach for creating practical open-domain conversational agents.