AI: Bite by bite

Presentation slides from the AI congress, London, January 2018

Transcript

1. AI Bite by bite Dejan Petelin Head of Data Science
2. BSc + MSc, Computer Science Computer science with focus on algorithms. Internship at the national institute. 2004 2009 PhD, Machine Learning On-line learning of Bayesian kernel methods (Gaussian processes) on streaming data. S2DS Intensive and project-based, five weeks data-science training. 2014 Gousto Applying AI (machine learning and optimisation algorithms) to disrupt the online grocery market. Dejan Petelin Head of Data Science
3. Simple model Leading proposition Fueling family life through good food Use AICONGRESS to get 30% off for the first 3 boxes! • You pick from 24 recipes each week. • We deliver a box of wholesome ingredients in exact proportions with step-by-step recipe cards. • No planning, no supermarkets and no food waste! • Most choice (24 is just the start). • Most delivery options (7 days, with am & pm slots) • Best price!
4. Exponential growth is fun, but also very challenging. Especially, when you are pushing boundaries in customer propositions at the same time. And of course growing the team and keeping your bar high. Exciting journey so far 2018
5. Irene Data Scientist Marc Data Scientist Manuel Machine Learning Eng. Catherine Data Scientist Team Diverse team - gender and background wise - to cover different aspects and views. Centralised team - part of the tech team, working on various cross-functional projects.
6. How we do AI (data science) at Gousto How we doubled the throughput of our automated picking line
7. How we do AI (data science) at Gousto
8. 1 Create focus Create a separate, dedicated data analytics function, all data insight queries go there. 2 Build what customers need Use agile product management and scrum to deliver value early. 3 Create production ready products You build it, you run it. 4 Measure success Make it accountable to the investment being made. Building innovative data products
9. Collect Everything Store Everything Expose Everything * * Whilst maintaining data security apps web microservices 3rd party airflow (ETL) amazon Redshift (data warehouse) amazon S3 (data archive) (unified log) data scientists data products business users periscope (analytics and dashboards) Amazon DMS (data migration)
10. Demand Planning Personalised Choice Warehouse Optimisation historic orders data Stock Handling X total orders recipe A 3.4% recipe B 2.8% recipe C 4.1% … onions chicken garlic ... 10,000 8,000 7,000 ... orders routes 10012 10012 10012 10013 10013 ... S001 S004 S020 S005 S013 ... R1 R2 R3 R1 R2 R3 0.8 0.8 0.6 0.6 0.4 0.4 1.0 1.0 1.0 R2 onion R3 beef chicken parsley R1 defaults choosers station picks
11. How we doubled the throughput of our automated picking line
12. Automated picking line Before the automation we were picking ingredients manually, i.e. ingredients were placed on tables, grouped by recipes. The automated picking line allows us to be more flexible in terms of customer propositions, yet efficient enough to make them viable.
13. Specific setup Gousto • Fresh items • Product catalogue changing on a weekly basis • Item volumes changing on a daily basis Gousto’s setup • Items are placed on different locations every week (weekly redesign of our warehouse) • Optimise pickface AND finding shortest path to collect items (difficult to predict due to completely new setup) Typical e-commerce • Items with a long shelf-life • Product catalogue are changing rarely • Item volumes are cyclical (seasonality) Typical setup • Items are placed on the same locations (no need to redesign the warehouse) • Finding the shortest path to collect all items (can be predictable due to historic data)
14. Trillions of trillions of combinations Thousands of slots ... hundreds of items ... and numerous rules. p a l e t fi t ... c i k r a n g s o l t be c lo e c he ... it n is t e p d o y t e h d a … in d e s l i l ti , s o d a c as s e ...
15. Genetic algorithms Well known optimisation algorithm It belongs to the larger class of the evolutionary algorithms, mimicking the same mechanism used by mother nature in evolution by natural selection, i.e. mutation, crossover & selection. Suitable for solving binary problems Like - in our case - placing items on slots, where a lot of other optimisation algorithms struggle. Recently got a lot of attention in Deep Learning Uber’s AI lab published a paper Deep Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning.
16. Production deployment Critical part of the business • There is a running time limit • If delayed or it fails, we are in danger of not being able to deliver boxes to our customers Not after the ‘perfect’ solution • Parallelisation / batching to speed it up • Trade-off between batch size, number of generations and performance of the end output forecast recipe SKUs pickface optimisation pickface (WMS) orders (mon) orders (tue) orders (wed) ... ... ... ………….
17. Production deployment Critical part of the business • There is a running time limit • If delayed or it fails, we are in danger of not being able to deliver boxes to our customers Not after the ‘perfect’ solution • Parallelisation / batching to speed it up • Trade-off between batch size, number of generations and performance of the end output forecast recipe SKUs Monitor and react! pickface optimisation pickface (WMS) orders (mon) orders (tue) orders (wed) ... ... ... ………….
18. Double the throughput Automated warehouse design Minimised manual work and avoided a lot bottlenecks - comparing to the initial plan to set up warehouse manually (by recipes). Overriding orders routing Significant reduction of station visits, which lead to 90% increase in the throughput, i.e. packed boxes per hour. Capacity / capability increase It directly improved our margin, but it also allowed us to keep growing and to invest into new factory later. 1 2 3 apr may jun jul aug sep oct nov dec
19. Thank You (and we’re hiring) twitter: @GoustoTech blog: techbrunch.gousto.co.uk apply: gousto.co.uk/jobs