 So our next presentation is by Millie Chapman of the University of California, Berkeley, talking to the role of AI in post 2020 biodiversity targets and effective management something very close to the work we're doing here at. Hi, my name is Millie Chapman and I'm a PhD candidate at UC Berkeley and the Department of Environmental Science Policy and Management. Today I'll be presenting about the use of AI in conservation decision making. I'll highlight some of our recent work showing that reinforcement learning holds promise to improve our capacity to effectively make decisions in the wake of global environmental changes. I'll also present some of our work on the ethical issues and power considerations that are critical to equitable and effective application of algorithms to conservation decision making. While human impacts on ecosystems rapidly accelerate our techniques for observing those impacts are increasingly outpacing our capacity to respond to them. Advances in artificial intelligence have largely been applied in conservation science to more precisely document the impacts of environmental change through improved and automated data classification workflows. But solutions to the global biodiversity crisis will rely on more than our capacity to more precisely estimate animal populations or dynamics. We must also improve our capacity to translate imperfect information about environmental systems into effective management decisions. So throughout this presentation, I'm really going to focus on the work that our group is doing to explore applications of algorithms and decision making and decision support processes. But can AI really help us make better conservation decisions? That question is inherently both technical and social, and the two parts of this issue are interdependent. The few projects I'm going to briefly present on today are addressing this larger question and the result of collaborations between environmental scientists, computer scientists, sociologists, and ethicists. I'm going to start today with some of the technical work happening in our group, which is largely exploring deep reinforcement learning as a method for addressing environmental issues. Quickly before I get started, I just want to explain what reinforcement learning is. It's really just an area of machine learning concerned with how intelligent agents ought to interact with an environment in order to maximize some notion of cumulative rewards. In the interest of time, I'll spare you all the details here, but if you're interested in learning more, I'd encourage you to look at our lab's recent preprint on these ideas. But in the most basic terms, reinforcement learning consists of an environment, which might be a fish dock or a wildfire or any number of environmental problems. And then there's an agent. So the agent takes an action in the environment. So in the case of fisheries, the agent might choose a harvest quota. And then the agent receives an observation of the environmental response to its action and a reward. So over thousands or millions of iterations, the agent learns how to effectively interact with the environment to maximize its rewards. So some of the advantages of using reinforcement learning over traditional optimization or decision theoretic approaches include flexibility and problem formulation. The environment can be really indefinitely complex and have large state spaces. Now the agent might have trouble in a particularly complex environment. But there's no true constraints on the way in the same way that there is with traditional approaches to control problems. Additionally, reinforcement learning can leverage existing ecological models and simulations for agents to interact with. In this preprint, we show that reinforcement learning holds promise for informing decisions for a number of environmental crises using primarily simple examples. So kind of what next. We're continuing to work to develop more complex case studies, for example, can reinforcement learning help us design dynamic marine reserves in a changing climate. And second, we're continuing to work to develop methods and theory around this approach. So what conservation questions are most suitable for reinforcement learning and what reinforcement learning algorithms are most effective at solving different conservation problems. So let's go back to the second main part of effectively leveraging AI and decision making processes, which is that these are not just technical questions about algorithm design. But they're also social and ethical questions for whom are these algorithms really creating better decisions. In this paper, we explored these broad questions in the context of algorithmic recommendations high seas conservation. Among other points we show that because algorithmic approaches at a global scale, largely consider the costs and benefits of conservation action to only a subset of human actors. Their applications can stand to reinforce power asymmetries due to do partially to data disparities. For example, sea bed mining and sea bed mining is largely done by industrial countries. So if we're trying to maximize biodiversity protection on the high seas while minimizing the impact on some sort of economic profit, we are largely considering large industries and largely ignoring coastal communities. Moreover, algorithms used in decision making processes have the potential to shift power dynamics along multiple axes. Maybe most obviously algorithms can shift who pays the cost for conservation and who receives the benefits. But algorithmic approaches to conservation decision making can also shift who decides what questions we ask and how we frame those questions. So is there an equitable path forward, and can we harness these technologies in a way that benefit more than just the powerful few that are designing and implementing them. This question continues to be an active part of our research, but in our, in a recent paper, where we're looking specifically at high seas conservation we argue that algorithms can be a part of an equitable decision making process. The interventions are taken throughout each stage of the prioritization process from procedural justice in the funding and team assembly for a project to the integration of social science data and methods in algorithm to distribution equity assessments of algorithmic solutions to ensure that optimal solutions are also equitable solutions. And with that, I'd like to acknowledge and give a huge thanks to the team who led a lot of the work I presented today, and to the many other collaborators who made this work possible. I'd also like to thank our funders for their support. And thank you for listening and please not hesitate to reach out with any questions. Thank you very much Millie thanks for sharing your story. In FAO, we do have a vision for more productive and more sustainable fisheries and agriculture, especially from the fisheries and agriculture division section of the agriculture outlook. But I think more and more people are recognizing that the word equity or equitable needs to be part of those types of visions because we've seen too many problems where we're trying to deliver outcomes. And my question to you is, when you've run these more theoretical overviews, where do you, where do you see the biggest opportunities? Is it maybe maybe I will just, when you started, did you have any surprises on what you found when you started to really do the research on it or did anything come up again and again, which just showed you that that's so important. Thank you. Yeah, I think that's a really interesting question and I think it's specifically an interesting question when we start to think about the use of these algorithms on the global scale. A lot of the data we have at a global scale, particularly on human dimensions is incredibly bias. It's very biased towards large industry, largely because that's where we have sensing technologies on vessels or whatever else it is. And it's largely not in human dimensions of coastal communities or other communities that might be underrepresented on a negotiation table on a global scale. And so I think it's really interesting to start to think a little bit more about how we start to mitigate some of those data disparities that exist in human dimensions data. And I think that there are also data disparities in ecological or environmental data. Without a doubt, we have more data in parts of the world than other parts of the world. But I think it's, I've found kind of through some of the research we've been doing over the last couple of years that that's particularly striking when we're thinking about human dimensions. Thank you, Mili. Matt, can I turn to you? Yeah, fantastic presentation. Really interesting and very important to explore this. Can you tell me more detail about how ecological simulations can be integrated with reinforcement learning methods and other resources for making this possible? Yeah, that's a great question. And thanks again for having me here today. This is a really cool form. Yeah, so there are a lot of cool resources in terms of like applying reinforcement learning to different environments, which I kind of presented in that one slide about some of the technical work that our lab is doing. And the preprint that I referenced has some great resources of how you develop environments to test different reinforcement learning algorithms on. And yeah, there's a whole kind of project being run largely by computer scientists for testing different reinforcement algorithms. It's called AI Gym. And you can develop our environments of your own. And so we've been developing environments that kind of mirror environmental problems that are hard to solve with traditional optimization approaches. And yeah, kind of seeing which reinforcement learning algorithms are effective in those different environments. But yeah, I guess I would point people towards that preprint and or kind of AI Gym and some of the resources that exists there. And do you integrate any sort of behavioral change psychology into your work as well? Yeah, I think that's a really interesting question. So a lot of this reinforcement learning work is more like how do we best like control a system or how do we like best make decisions about a system and less kind of about understanding the dynamics of a social environmental system that might be experiencing what policy changes and behavioral changes. But I think that's a really interesting point that you raised that a lot of these methods are also interesting for starting to understand the dynamics of human decision in a complex environmental system. So if you have an agent interacting dynamically with a system, you could imagine kind of treating that agent and kind of exploring how behavioral changes might influence the trajectory of an environmental system.