Machine learning applied to food delivery logistics

• Grant number: 17/15401-1
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: May 01, 2018
• End date: January 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Adauton Machado Heringer
• Grantee: Adauton Machado Heringer
• Company: Cozo Tecnologia Ltda
• CNAE: Atividades profissionais, científicas e técnicas não especificadas anteriormente
• City: São Paulo
• Associated researchers: Flávio Soares Corrêa da Silva
• Associated research grant(s): 18/22516-2 - Machine learning applied to food delivery logistics, AP.PIPE
• Associated scholarship(s): 18/11827-7 - Machine learning applied to food delivery logistics, BP.TT; 18/12107-8 - Machine learning applied to food delivery logistics, BP.TT; 18/11676-9 - Machine learning applied to food delivery logistics, BP.PIPE

Abstract

After reaching R$ 1.25 billion in 2015, the Brazilian food delivery market through online apps is expected to reach R$ 12 billion by 2021. In spite of this significant growth, the majority of restaurants will not get the minimum scale to justify investments in both fleet and employees, causing a natural selection of the larger and more expansive ones. Therefore, average price of delivery food is twice as much as average food price in the establishment, preventing the participation of classes C and D in this market. As a solution, Cozo develops logistic intelligence software that connects restaurants and independent motorcycle carriers to end costumers, enabling delivery service without fixed investments by the restaurants. As differentiation, Cozo uses machine learning techniques to optimize the general logistic system, benefiting the final customer with speedy delivery, the motorcycle carrier with cost reduction, and the restaurants with a new online food court. In the context of machine learning, the research project is centered on the application of Bayesian Additive Regression Trees to predict orders' preparation time. Such modeling will allow a complete inference of the prep time a posteriori, via Markov Chain Monte Carlo sampling. Therefore, not only the average prep time might be estimated but credible intervals and marginal effects might be determined. Besides precision on order's prep time estimation, it is expected that the algorithm be able to learn and refine its prediction accordingly to restaurants' structural changes such as amount of orders, time scheduling, and staff changes. A positive result in this research project along with advanced routing techniques will enable the disruption of online food delivery market through entrance of cheaper restaurants. The main impact will be the reduction of average price from current R$ 60 to R$ 30, promoting inclusion of classes C and D in the context of smarter cities. (AU)