Felipe Montealegre-Mora

Profile

Hello there!

I come from a physics background and currently I am a postdoc on computational ecology at UC Berkeley in Carl Boettiger’s lab.

My B.Sc. was in my home university, Universidad de Costa Rica, while for grad school I moved to Cologne, Germany. There, I worked on quantum information and representation theory in the group of David Gross. In 2022 I graduated and moved to Berkeley and here I am!

My research summary

My research focus spans several applications of machine learning and data science methods to ecology. Here is a short summary of the main points I touch upon:

• Reinforcement learning (RL) for dynamic ecological management.
  • RL for fishery management: For fishery ecosystems with complex dynamics, the RL tools we develop produce adaptive policy recommendations on fishing quotas. Right now, we are developing these tools for the management of highly stochastic populations with spasmodic recruitment events—fisheries for which infrequent recruitment booms are the main driver of the population dynamics. Our current focus is on Walleye fisheries.
  • RL for caribou conservation: Caribou populations have been declining in Canada due to an increased mobility of wolves resulting from anthropogenic causes. This project seeks to find policy recommendations for caribou conservation in the form of e.g. targeted habitat restoration, wolf culls and moose culls (an alternative nutrition source for wolves).
  • RL for invasive green crab management: Green crabs are an invasive species that has been particularly detrimental for, among others, shellfisheries in the western American coast. This project seeks to optimize temporal resource allocation to mitigate green crab colonization in Bays and estuaries.
  • Curriculum learning: curriculum learning approaches to ecological management problems with model uncertainty.
• Ecological time-series forecasting
  • NEON forecasting: starting to ramp up coding to submit to the ecoforecasting challenge! My focus is on employing machine learning forecasting tools to improve forecasts.
• Combining geospatial and ecological data for environmental justice
  • This is my newest project! Here we use open-source GIS tooling in R and python in order to reproduce classic results from enviornmental data science.
  • My jump-off point is this example in which we reproduce the result that redlining impacts access to green spaces by associating historical HOLC grades with NDVI from satellite images in San Francisco.
  • Currently I am working on a tutorial for quantifying species richness in protected areas as a function of management type. In particular, this tutorial seeks to reproduce the result that indigenous-management of protected lands is at least as beneficial for ecosystems as western management in the form of national parks (and sometimes more so). A first angle we’re exploring to quantify this comparison is via measures of species richness.

My research soapbox

I am a physicist by training, beyond any doubt—my entire education history is physics-centred, where I have been passionate about different topics over time: experimental solid state as an undergrad, theoretical condensed matter and quantum information as a masters student, and finally converging on representation theory applied to quantum computing as a PhD student. The question which might come up at this point is: and how did you come to work at an environmental science, policy and management department?

This was a conscious decision to switch my research focus towards topics which I feel are of utmost importance nowadays. It came as an answer to the question how do I best use the skills I have learned in my academic upbringing? During the last period of my PhD, particularly after 2020, I began asking myself this and questioning whether I should veer my research towards more applied topics. This questioning lead me to my current position, where I seek to apply state of the art mathematical and computational techniques towards solving environmental problems.

Soapboxing on the RL project

I am currently very excited about the possibility to leverage reinforcement learning (RL) algorithms in complex ecological management scenarios. We recently published a paper on fishery management showing how reinforcement learning can yield highly dynamic policies that have higher long-term productivity than well-established management strategies (e.g. MSY, constant escapement).

This initial paper, while very encouraging, trained RL algorithms on rather stylized mathematical models and served as a proof of principle. From here, I have been engaged in several collaborations to bring algorithmic ideas from the RL world into the ‘real world’ of environmental management. Here, my guiding principle is: how can we make the most out of these tools, while understanding their limitations? How could these tools be realistically incorporated into environmental decision making?

Other RL projects in the making

Curriculum RL. (repo) Mathematical models are useful lies and ecological management has to account for this. This fact made me go deep into the world of curriculum learning, an approach to RL in which the algorithm learns to interact with a variety of environments and perform well in all of them. I have written a series of tools in the python package EcoCuRL for this exact purpose. As a sample application, I train an agent on a fishery management problem with uncertainty in the intrinsic growth rate of the fish population.

RL Ecological Benchmarks. (repo) I’m pursuing is on benchmarking RL algorithms on ecological probelms. Benchmarks on classic OpenAI environments are very helpful: whenever faced with an optimal control problem, these benchmarks help us decide how to choose and tweak an RL algorithm to fit our problem best. Because using RL in ecology is rather new, benchmarks for RL performance on these types of problems are few and far in between. Because of this, using RL on ecology still requires a lot of tinkering before the algorithm produces good results.

I hope to help close this gap with this work. I will provide a series of benchmark population dynamics problems of varying complexity. This can guide future researchers and managers when choosing, e.g., the optimization algorithm and the hyperparameters to use in their RL management approaches.

More

Check my GitHub profile for up-to-date code on these and other projects!

My PhD research

Throughout grad school I worked at the interface of representation theory and quantum computing. My work spanned abstract algebra, numerical linear algebra, algorithm development and coding. I made contributions characterizing Clifford tensor-power representations, approximate unitary t-designs, measures of non-stabilizerness, and developing algorithms for numerical representation theory.

Outreach

I was an active member of the science communication collective ManyBodyPhysics. There I produced a cute little article on randomized benchmarking for quantum computers which I am very proud of.

Contact

Don’t hesitate to reach out to me at felimomo [at] berkeley.edu
Media: LinkedIn, GitHub, Google Scholar, ManyBodyPhysics

My office is Wellman 210, at the UC Berkeley campus.

Get new content delivered directly to your inbox.