Felipe Maldonado

Field of Expertise

My name is Felipe Maldonado, I am a Lecturer (Assistant Professor) at the School of Mathematics, Statistics, and Actuarial Science, at the University of Essex. Before that I was a Postdoctoral Researcher at the Technical University of Munich (Department of Informatics) working with Martin Bichler on electricity market design.

Before moving to Germany, I did my PhD studies at the Australian National University (School of Computer Science), working with Pascal Van Hentenryck and Gerardo Berbeglia. During my stay in Australia I was also part of CSIRO-Data61. Prior to that I worked for 2 years at the University of Concepcion as a fixed-term lecturer. Before that I studied Mathematical Engineering at the University of Chile, and I worked with Roberto Cominetti on my final thesis about myopic agents with adaptive behaviour.

My expertise lies between optimisation and analytics, where I have developed research and practical projects in a series of fields such as

• Reinforcement Learning
• Energy Analytics and Market Design
• Game Theory
• Operational Research
• Logistics and Supply Chain
• Consumer Behaviour and Revenue Management
• Fintech

Expertise

The research I have been developing over the years is inspired by challenges coming from several industries (e.g, Energy, Supply Chain and Logistics), and quantifying what is a good-enough solution, is a question that has resonated throughout my career. I constantly consider the trade-off between how efficiently a solution can be computed, and how close to optimal it is. The use of data has played a key role on this. The efficiency/optimality trade-off becomes critical when developing models for industry-related problems, this is because a very complex model that captures more accurately the reality, could have some drawbacks regarding computability. For instance, during my PhD studies at the Australian National University, my research focused on developing discrete choice models that represented consumer behaviour in online markets, where some of the key features were the order in which products are displayed and the effects of previous purchases.

To have a clear idea of what are the real implications of my research, I have maintained a continuous involvement with industries over the years. After graduating from Mathematical Engineering in 2012, I joined a consulting company where we worked solving problems such as job scheduling in a wine bottling process. When I moved to Australia to start my PhD I was funded by CSIRO-Data61, that also gave me the opportunity to work on applied projects. Particularly I worked as a data analyst developing a new billing schedule for Queensland Urban Utilities, minimising the impact over the consumers (potential spikes on the predicted bills). In 2020, I was invited for a podcast session organised by an Australian tech company Hivery, titled Retail Mavericks. Hivery develops artificial intelligence tools applied to the retail industry, and in our podcast (https://www.hivery.com/podcasts/dr-felipe-maldonado) we talk about market design, optimisation and their application to problems such as shelf optimisation.

Before joining the University of Essex, I was a postdoctoral researcher at the Technical University of Munich (TUM) working on new proposals for electricity market design. In this position I collaborated with a multidisciplinary team as part of the Kopernikus-Projekt SynErgie, which was composed by members from several German universities and research centres, as well as industry partners such as transmission system operators in Europe. In this collaboration we have highlighted the main issues of the current electricity spot market design in Europe, which by the time were not as obvious as they are now. We also provided recommendations such as moving to a locational marginal price-based system along side with more descriptive bidding formats, and proposed new pricing mechanisms.

My research roughly can be divided in three research lines:

RL1: Energy Analytics and Market Design This research line aims to address the ubiquitousness of non-convex costs in markets (e.g. start-up costs, minimum production, set-up times). Those non-convexities make optimisation problems much harder to tackle, especially when they need to be solved very frequently. For example, in the context of intraday electricity markets, large mixed integer linear programs are solved every 5-15 minutes, providing both an optimal allocation and clearing prices. Non-convexities present a fundamental challenge for pricing mechanisms that aim to provide an efficient allocation in competitive markets. A poor market design can send the incorrect investment signals, leading to inefficiencies. We have been experiencing this in our recent energy crises, prices in the wholesale market have reached historical records given that they are set up by the most expensive resource (through the merit-order system and marginal pricing). Different countries have taken palliative measures such as implementing revenue caps, but it is uncertain how much these measures will affect the long-term investment signals, which are critical for the Net-Zero goals.

My research in this area has been heavily motivated by electricity markets, but the techniques (e.g. mixed integer linear programming with non-convex objective functions) can be easily extended to many other settings: for instance logistics with flexible demand (e.g., products to be delivered over three days), or cloud computing tasks allocation. In this area I have been also exploring the use of Reinforcement Learning (RL) as a way to learn optimal policies for local energy markets, where the agents have as actions, tariffs for their clients aiming to achieve desirable levels of demand flexibility.

RL2: Logistics and Warehouse Operations/Management This research line was born through previous collaborations with Siemens (co-supervision of Master student at TUM). Our first project was about Just-in-Time Manufacturing, where we aimed to reduce inventory costs in production processes. Our methodology used historical and real-time data to optimise their operations. We developed a model for prediction of lead-times, helping to monitor the suppliers’ performance; and a model for demand forecasting of different materials to identify slow/fast-moving items. Determining critical levels where actions were needed. We used different machine learning and time-series algorithms, leveraging the computational capabilities of cloud services (AWS). Due to non-disclosure agreements, we did not publish on this topic and it just remained as an applied project.

At Essex, I have been developing two projects that follow a similar structure, but in a different domain. The first one consisting on a multi-vehicle consolidation system for last-mile logistics, where van drivers can share their vehicle loads with smaller vehicles (e.g., cargo bikes) in order to reduce their impact to congestion and pollution in busy areas. I have used data from Uber-Movement (routes distances are weighted by congestion) to decide upon an optimal policy (through the use of RL and Monte Carlo Tree Search methods). The second project has recently started and it is associated to an order batching and picker routing problem for robots working on a retail warehouse. One of the main challenge here is to deal with multiple robots that carry out different orders (affecting the routes of incoming robots).

RL3: Text Analytics and Anomaly Detection A second collaboration with Siemens consisted on CT scan machines preventive maintenance. In there we used log-data for those machines to create classifiers based on anomaly detection and explainable analytics. We identified critical time windows where an anomaly might occur, and preventive maintenance was recommended.

Through KTP projects at the University of Essex I have been working on several text analytics’ developments, for instance creating recommender systems based on large language models using Insure- tech data. Another project is about the creation of a language model-based tool to understand the mismatch between Standard Industrial Classification (SIC) codes and what UK companies actually do. This can help for instance, to have a more systematic way to obtain insights on investment opportunities and access to special funds (e.g., Covid-19 or Net-Zero associated grants).