JOIE: A Collaborative Multi-agent Reinforcement Learning Framework for Dialog Action Decomposition

Research Paper: JOIE: A Collaborative Multi-agent Reinforcement Learning Approach

Overview

JOIE (Joint Operation Inference Engine) is an innovative framework designed for dialog action decomposition using a collaborative multi-agent reinforcement learning approach. This repository contains the implementation of the JOIE-3 algorithm, a cornerstone of our research.

Key Components

Core Algorithm

• The main implementation of JOIE-3 is located in /JOIE Main/JOIE.py.

Configuration and Parameters

• For parameter settings, refer to demo.json.
• Action decoder configuration can be found in /JOIE Main/multiwoz_vocab_action_decoder_hrl.py.
• Memory settings are detailed in the ReplayHR3 class within /JOIE Main/replay.py.

Prerequisites

JOIE is built upon the ConvLab platform. To run JOIE:

1. Install ConvLab: Visit ConvLab GitHub Repository for installation instructions.
2. After installing ConvLab, use the following command to execute the JOIE framework:

python run.py demo.json rule_wdqn_Co3Jo train