Background/Question/Methods Advances in both available data and computing power have begun to open the door to a greater role for machine learning (ML) in addressing some of our planet's most pressing environmental problems, such as the growing frequency and intensity of wildfire, over-exploited fisheries, declining biodiversity, and zoonotic pandemics. But will such approaches really help us tackle a changing planet? The reliance of dominant machine learning approaches on existing data can make them a valuable tool for automating well-understood steps in an analysis, but may also make them a poor tool to predict the long-term dynamics of a no-analogue future and the non-representative distribution of available ecological data. Supervised Learning methods have proven very successful in tasks such as image classification, with the promise of being able to automate species identification from cameras, acoustic sensors or satellite images. Unsupervised learning methods hold out a more ambitious promise of identifying features or making predictions without first requiring a large training set labeled data. The least attention has been paid to what may be the most promising member of the ML triumvirate: Reinforcement Learning (RL). RL trains a software agent to maximize some objective -- notable examples include navigation in autonomous robots or Google's AlphaZero. RL requires agents to make decisions in dynamic and uncertain environments that plan ahead and change the state of the board. Also unlike other approaches, such agents are typically trained, not on mountains of data alone, but on simulations.
Results/Conclusions I describe how this creates a bridge between the rich, mechanistic or process based models of our field and the power of ML to discover creative solutions to complex decision problems. Rather than seeking to displace process-based models of the past century with opaque machines, I illustrate how we can use those models to drive realistic simulations which we can then train leading RL algorithms to manage. I illustrate this approach with examples from fisheries management, ecosystem tipping points, and wildfire spread. These examples are only a proof-of-principle which illuminate not only the promise, but also some of the pitfalls ahead. The pitfalls are not only technical, but include issues of ethics and power, particularly if the algorithms or data are proprietary. I conclude with a discussion of how an open, transparent and reproducible approach can help mitigate some concerns, while also offering a more effective interface between teams of researchers from both ecological and computer sciences.