Multi-Agent Reinforcement Learning-Based Real-Time Cooking Task Scheduling Optimization for Multi-Chef Collaborative Cooking

Nowadays, with the increasing living standard of people, there is a growing demand for the catering industry. Cooking efficiency is an important factor for restaurants to be highly competitive, and the efficient scheduling of cooking tasks of chefs matters the most. To enable efficient multi-chef collaborative cooking for restaurants, in this paper, a novel real-Time cooking task scheduling method, MAPPO-LSTM, is proposed. In the MAPPO-LSTM, firstly, the proximal policy optimization (PPO) algorithm is augmented with a centralized training and distributed execution scheme to address the modeling of cooking task allocation for multiple chefs. Besides, convolutional neural networks (CNN) and long short-Term memory (LSTM) are introduced to the actor network to mine temporal features of the cooking environment and to enhance the memory of historical cooking behavior sequences, respectively. Experiments are conducted using the 'Overcooked' video game as the simulation environment. Compared with benchmarking methods, the high efficiency of the proposed MAPPO-LSTM for the real-Time task scheduling in collaborative cooking is validated based on four indicators.