Reinforcement Learning for Conservation
Can Deep Reinforcement Learning Improve Conservation Outcomes?
Open Benchmarks for Conservation Agents
Deep reinforcement learning (RL) has proven remarkably successful at solving some difficult problems. However, so far most applications have focused on problems based on games or robotics (e.g. pictured above: Atari moonlander, a standard benchmark in OpenAI Gym, and the Robel D'Claw, a Gym-compliant, trainable robotic claw developed by UC Berkeley researchers). Public benchmarks and open source implementations of RL problems help research teams to compare results and spur innovation. Many conservation decision-making problems involve elements which make them particularly challenging, such as complex, nonlinear dynamics, tipping points, stochasticity. This project seeks to engage researchers in both Artificial Intelligence (AI) and conservation biology to develop realistic conservation environments and to train, test, and compare performance of the RL agents on those environments.
Environments
Forest Fires
How can we best respond to the growing magnitude and frequency forest fires around the globe?
alpha version
Overfishing
Decades of overfishing have left many fish stocks depleted worldwide. Today, effects of global warming may frustrate rebuilding of even well-managed stocks.
beta version
Endangered Species
Should conservation efforts prioritize hotsots of species diversity, or land that is most threatened? Focus on the species at risk or most likely to benefit? There are no easy answers.
alpha version
Climate Change
How do we balance mitigation and adaption strategies to rising sea levels?
alpha version
Bark Beetle Devastation
Fueled by warming temperatures, bark beetles are destroying tens of thousands of square miles of forests across North America, Europe, and Siberia.
Novel zoonotic pandemics
How best do we respond to emerging pandemics such as COVID-19?
alpha versionAgents
This section will provide links to code and documentation for trained agents on particular challenges. These examples will provide a convenient starting place for other researchers seeking to improve the performance of RL algorithms on each of the challenges included here.
Stable Baselines V3
Examples which use PyTorch-based stable-baselines3 implementations existing published algorithms on each of the gyms
Leaderboard
This sections will track the performance of all known AI agents on each of the challenge environments. Submissions from all contestants are welcome. We only request that you share code to an open-source implementation of your agent that can be run against the benchmark environments.
