CompSustNet is a network of researchers who are developing new computational approaches to addressing challenges in environmental and societal sustainability. In my role as CompSustNet’s Director of Outreach, Education, Diversity, and Synthesis, I plan to blog regularly about each of the topics indicated in my title, as well as posting publicly-accessible overviews of the research being undertaken in the network. For this latter area of responsibility, in particular, I will also recruit graduate and undergraduate students with an interest in communicating science and technology to the public, as fellow bloggers and co-authors of articles. Contact me if you are interested.
Because registration at most schools for Fall 2016 courses is starting soon, or is very recently underway, higher education in computational sustainability is a timely topic.
In the last 5 years or so there have been a number of college-level courses in computational sustainability, some with an emphasis in one subarea, but most with broader coverage. Here is a list of the courses that I know.
Sustainability and Assistive Computing (Bryn Mawr, Fall 2010). Broad coverage, with a large proportion of course dedicated to computing for societally important topics other than sustainability per se.
Computing and the Environment (Vanderbilt, Spring 2011). Broad coverage, for upper-division CS students, with programming prerequisites and some mathematical sophistication.
Topics in Computational Sustainability (Stanford, Spring 2016). A broad coverage of computational sustainability topics, with no formal prerequisites, but some mathematical and programming sophistication required.
Beyond the topics themselves, at least two things strike me as notable about these courses.
Instructors openly use and acknowledge educational content from other courses and other instructors, in addition to making nontrivial use of guest lecturers. As I plan for my Fall 2016 special topics course, entitled Computing, the Environment, and Energy, I will be using much from my earlier 2011 course, but I will also undoubtedly take Stefano Ermon (Stanford) and Dan Sheldon (UMass) up on their offers to share resources. I’ll consider drawing from other courses as well, and I expect to line up a few guest lectures over video conferencing as well.
These are not courses that fit the stereotype of “computing and society” or “computing and ethics” that require little or no computing background. Such introductory courses would be a good addition to the course portfolio, but consistent with the CompSustNet mission, the courses so far have opted to advance the education of committed and knowledgeable computer science students.
In the future, I will be posting more on computing courses, including my own Fall 2016 course. I invite others to contribute one or more posts on their experiences too. In addition to standalone courses in computational sustainability, I will be posting on “deep infusion of sustainability material into NON-sustainability-themed courses”. Deep infusion doesn’t seek to spinoff computational sustainability to specialized courses, but seeks to make it ubiquitous across the entire computing curriculum.
Douglas H. Fisher is CompSustNet’s Director of Outreach, Education, Diversity, and Synthesis. The opinions expressed herein are Doug’s and not necessarily those of Cornell University. Contact Doug at firstname.lastname@example.org.
Poverty, saving species, repowering the world with renewable energy, lifting people up to live better lives—there are no easy answers to guiding our planet on the path toward sustainability. Complex problems require sophisticated solutions. They involve intricacy beyond human capabilities, the kind of big-data processing and analysis that only advanced large-scale computing can provide. CompSustNet, funded by an NSF Expedition in Computing award, is a vast research network powered by the nation’s top university computer science programs, charged with applying the emerging field of computational sustainability to solving the world’s seemingly unsolvable resource problems. Put simply, the project enlists some of the top talents in computing, social science, conservation, physics, materials science, and engineering to unlock sustainable solutions that safeguard our planet’s future.
Computational Sustainability is, at its core, the belief that with sufficiently advanced computational techniques, we can devise sustainable solutions that help meet the environmental, societal, and economic needs of today while providing for future generations. In much the same way IBM’s supercomputer Watson could defeat any challenger in Jeopardy!, computational sustainability posits that a computer-engineered solution can be applied to virtually any of the world’s problems—from helping farmers and herders in Africa survive severe droughts to developing a smart power grid fueled entirely by renewable energy. CompSustNet is a large national and international multi-institutional research network led by Cornell University and including 11 other US academic institutions: Bowdoin, Caltech, CMU, Georgia Tech, Howard University, Oregon State, Princeton, Stanford, UMass, University of South California, and Vanderbilt University, as well as collaborations with several international universities. But CompSustNet is not just an ivory-tower enterprise, as it also includes key governmental and non-governmental organizations that specialize in conservation, poverty mitigation, and renewable energy, such as The Nature Conservancy, The World Wildlife Fund, The International Livestock Research Institute, The Trans-African Hydro-Meteorological Observatory, and the National Institute of Standards and Technology.
CompSustNet’s core mission is to significantly expand the horizons of computational sustainability and foster the advancement of state-of-the-art computer science to achieve the scale to tackle global problems. Research will focus on cross-cutting computational topics such as optimization, dynamical models, simulation, big data, machine learning, and citizen science, applied to sustainability challenges. For example, computational sustainability is being put to work to resolve the problem of providing wetlands for shorebirds that migrate from the Arctic through California during a time of drought. As California gets drier, the shorebirds have nowhere to stop, rest, and refuel by eating wetland invertebrates. Scientists are developing new dynamic precision conservation techniques that use complex, big-data models to tackle the problem with NASA satellite imagery, meteorological forecasts, and citizen science in the form of thousands of bird location sightings from the Cornell Lab of Ornithology’s eBird checklisting app for birdwatchers. Through partnership with The Nature Conservancy, the program forecasts when and where wetland habitat would be needed for shorebirds, and the Conservancy pays Central Valley rice farmers to flood their fields at opportune times—providing benefits for birds and farmers at a time when extreme drought is making life tough for both. In similar ways, computational sustainability projects will also be hard at work innovating automated monitoring networks to protect endangered elephant population from poachers, promoting the discovery of novel ways to harvest energy from sun light, and generating algorithms to manage the generation and storage of renewable energy in the power grid.
Advancements in computational sustainability will lead to novel, low-cost, high-efficiency strategies for saving endangered species, helping indigenous peoples improve their way of life, and scaling renewables up to meet 21st century energy demand. CompSustNet is like the seed, the venture capital, to help the field of computational sustainability achieve what’s possible.
The research leadership and contacts from all institutions are listed on the CompSustNet website.
CompSustNet is “a transformative computational sustainability network, bringing together computer scientists, environmental and social scientists, biologists, physicists and material scientists to expand the nascent field of computational sustainability,” says Carla Gomes, the Principle Investigator on the lead Cornell award.
This is the second Expeditions of Computing award led by Gomes on Computational Sustainability, a testament to how important it is that computer scientists and researchers in related disciplines take up the challenges of environmental and societal sustainability.
In addition to the funded network nodes referenced above, there are many other collaborating network members from the United States and across the world.
Commonwealth Scientific and Industrial Research Organization (CSIRO) – Land and Water Flagship
Data61 (formerly NICTA), Australia
Delft University of Technology, Trans-African Hydro-Meteorological Observatory (TAHMO)
Distributed Sun LLC
Earth Networks Inc.
International Livestock Research Institute (ILRI)
National Institute of Standards and Technology (NIST) – Materials Measurement Laboratory
Nature Trust of British Columbia
Northwest Watershed Research Center
Swiss Federal Institute of Technology in Zurich (ETHZ)
Swiss Federal Institute of Technology Lausanne (EPFL)
Technical University of Dortmund
Technical University of Lisbon
The Nature Conservancy (TNC)
United States Geological Survey – Patuxent Wildlife Research Center
Universidad San Francisco de Quito, Instituto de Ecología Aplicada (ECOLAP)
Università di Bologna
University College Cork, Insight Centre for Data Analytics
University of British Columbia
University of British Columbia
University of British Columbia,
University of Idaho,
University of Lleida – Artificial Intelligence Research Group
University of Maryland
University of Oxford
University of Sydney
University of Washington, Coastal Observation and Seabird Survey Team (COASST)
In the coming weeks, months, and years, look here and to the CompSustNet website, for updates on CompSustNet activities, as well as for opportunities for participation with the network. Please tell us about computing research in computational sustainability, whether previously called by that name or not, and please tell us of efforts to effectively communicate computational sustainability research to the public.
Douglas H. Fisher is the Director of Outreach, Education, Diversity, and Synthesis for CompSustNet, and is an associate professor of computer science at Vanderbilt University.
This week in Québec City the Twenty-Eighth AAAI Conference is being held and for the fourth year running there is a special track on Computational Sustainability. The range of talks look at ways to apply and extend ideas in Artificial Intelligence and Machine Learning for solving complex problems in Electric Vehicle Management, Bird Tracking, Disease Mapping, Power Management and Urban Emergency Preparedness.
You can see the full abstracts for the talks from the whole conference including the CompSust track online here. Also, if you subscribe or have access to AI Magazine, the magazine publication for AAAI, you can see some high level articles in the next two issues on the range and depth of CompSust research.
If you have research in this area that you want to submit then you’re in luck, due to AAAI changing it’s schedule from summer to winter the next AAAI CompSust deadline is just five weeks away! (Abstract:Sep10, Paper:Sep15) (Ok, so it’s a good thing if you have something ready to submit.) Good luck!
Happy New Year CompSust researchers! If you are in the Atlanta area take a look at this workshop being hosted by Georgia Tech. It looks like a good opportunity to build local awareness and community for CompSust problems and solutions.
Sustainability starts with the individual and extends to buildings, neighborhoods, cities, and regions. Workshop participants will simultaneously explore new approaches to achieving sustainable growth that can be enabled through innovations in computation, as well as the computational technologies themselves. Relevant computational technologies include data analytics, modeling and simulation, optimization, high performance computing, and distributed computing platforms. The event will include a brief overview to frame the discussion, a series of representative presentations covering ongoing research in sustainability challenges and computational technologies, breakout discussions to define overarching themes and synergies, and a networking reception to enable further discussion among researchers. While time does not permit everyone to give a formal presentation, discussion sessions are included to enable others to briefly discuss their research interests.
Computational Sustainability Workshop Date:Thursday, January 30, 2014
Location: Georgia Tech Hotel
Some news for researchers in Computational Sustainability updated throughout the month. If you know of relevant news for researchers working at the intersection of Computer Science and Sustainability share it with us in the comments, on Google+ in the CompSust community, or join the CompSust mailing list.
Workshop: SustainIT 2013 – The Third IFIP Conference on Sustainable Internet and ICT for Sustainability is an interdiscplinary conference between industry and academia to address topics both of the significant energy consumption of computing technology and as a potential actor in steering a more clever usage of energy resources.
Time and Place: October 30-31, Palermo, Italy.
Workshop: The First Workshop on Machine Learning for Environmental Sustainability (EnvSustML 2013) is collocated with the Asian Machine Learning Conference and aims to bring machine learning and data mining researchers and experts in the environmental domain together to tackle challenges towards environmental sustainability Deadline: Sept 30, 2013
Time and Place: November 13-15, 2013. Australian National University, Canberra, Australia
Knowledge of the distribution and ecological associations of a species is a crucial ingredient for successful conservation management, biodiversity and sustainability research. However, ecological systems are inherently complex, our ability to directly observe them has been limited, and the processes that affect the distributions of animals and plants operate at multiple spatial and temporal scales.
Very recently, large citizen science efforts such as eBird, a very successful crowdsourcing project by the Cornell Lab of Ornithology that engages citizen scientists and avid birders, is enabling for the first time world-wide observations of bird distributions. eBird has collected more then 100 millions of bird observations to date from as many as 100 thousand human volunteers and submissions (checklists) continue to grow with an exponential rate. With this wealth of evidence comes a plethora of challenges as the data collection and sampling designs are unstructured, follow human activities and concentrations, and are subject to observer and environmental biases. For example, sparsely populated states in the US, such as Iowa and Nevada, have very low frequency of observations whereas East and West Coast states have the highest continental counts. Furthermore, temporal variability and biases are also evident as annual submission rates peak during spring and fall migration which is the most exciting times for birders to observe multiple species.
In our AAAI paper, we propose adaptive spatiotemporal species distribution models that can exploit the uneven distribution of observations from such crowdsourcing projects and can accurately capture multiscale processes. The proposed exploratory models control for variability in the observation process and can learn ecological, environmental and climate associations that drive species distributions and migration patterns. We offer for the first time hemisphere-wide species distribution estimates of long-distance migrants (Barn Swallow, Blackpoll Warbler, and Black-throated Blue Warbler in Figure above), utilizing more then 2.25 million eBird checklists.
Until recently, most biodiversity monitoring programs that collect data have been national in scope, hindering ecological study and conservation planning for broadly distributed species. The ability to produce comprehensive year-round distribution estimates that span national borders will make it possible to better understand the ecological processes affecting the distributions of these species, assess their vulnerability to environmental perturbations such as those expected under climate change, and coordinate conservation activities.
D. Fink, T. Damoulas, J. Dave, (2013). Adaptive Spatio-Temporal Exploratory Models: Hemisphere-wide species distributions from massively crowdsourced eBird data. AAAI 2013, Washington, USA.
This is one of our series of posts on the latest research in Computational Sustainability being presented at conferences this summer. This time Theo Damoulas, a Research Associate in Computer Science, and a member of the Institute for Computational Sustainability, at Cornell University tells us about their new paper at AAAI in Washington this month.
The Twenty-Seventh AAAI Conference on Artificial Intelligence (AAAI-13) convenes next week in Bellevue, Washington USA. For the third consecutive year there will be a special track on Computational Sustainability, a nascent and growing field of computing that is concerned with the application of computer science principles, methods, and tools to problems of environmental and societal sustainability. This is not a one-way street, however, because sustainability problems force computer scientists into new theory, as well as new practice. For example, sustainability problems require extraordinary attention to solution robustness (e.g., so that a so-called optimal solution doesn’t catastrophically fail with an environmental change) and issues of uncertainty, ranging from uncertainties in environmental sensor readings to uncertainties in the budget awarded by a state legislative body for wildlife management!The 16 papers of the Computational Sustainability (CompSust) track of AAAI (http://www.aaai.org/Conferences/AAAI/2013/aaai13accepts.php#Sustainability) cover sustainability problems in natural environment, to include various forms of resource management (e.g., species management, wildfire control), and the built environment (e.g., smart grid, building energy usage). The CompSust presentations are arranged in four presentation sessions, all on Thursday, July 18, 2013. These sessions are organized by AI themes of MDPs and sequential processes, optimization and search, data mining, and multi agent systems.As in the past, the Computing Community Consortium (CCC) is graciously supporting best paper awards for the CompSust track, which will be announced at the opening ceremony on Tuesday, July 16.
If you have a very large decision making problem and want to find an approximately optimal policy, one of the best ways is often to use simulated trajectories of states, actions and utilities to learn the policy from experience.
In many natural resource management problems running simulations is very expensive because of the complex processes involved and because of spatial interactions across a landscape. This means we need an approximate planning algorithm for MDPs that minimizes the number of calls to the simulator. Our paper at AAAI presents an algorithm for doing that.
Example Problem : Invasive Species Management
One example of a natural resource management problem with this kind of challenge is management of invasive river plants. For example, Tamarisk is an invasive plant species originating from the Middle East which has invaded over 3 million acres of land in the Western United States. It outcompetes local plants, consumes water and deposits salt into the soil. This pushes out native grass species, fundamentally changes the chemistry soil and alters an ecosystem that many other species rely on (read more : studies of Tamarisk by NASA; how the Tamarisk Collective is removing Tamarisk and restoring riparian ecosystems). Dropped leaves also create a dry layer of fuel that increases the risk of fire in the already fire-prone West.
Seeds from plants can spread up or down the river network leading to a huge number of reachable states. There is a choice of treatment actions available in each part of the river: we can eradicate invading plants and/or reintroduce native ones. Each treatment action has a cost, but the more expensive treatments are more effective at supplanting the invading plants.
The Planning Problem
The planning problem is the following: Find the optimal policy for performing treatments spatially across a river network and over time in order to restore the native plant population and stay within a given budget level.
This problem can be represented as a Markov Decision Process (MDP) but it very quickly becomes intractable to solve optimally for larger problems. Ideally we want to find a policy with guarantees about how far it is from the optimal solution. PAC-MDP learning methods provide such guarantees by using long simulations to converge on a policy that is guaranteed to be within a given distance of the optimal policy with some probability (see Sidebar: What is an MDP? What does PAC-MDP mean?).
Most of the existing PAC-MDP methods look at a sequence of simulated actions and rewards and rely on revisiting states many times over and over to learn how to act optimally in those states. This does not fit the ecosystem management problem. In reality, we begin in a particular starting state S, in which the ecosystem is typically in some undesirable state far from its desired balance. The goal is to find a policy for moving to a world where Sdoes not occur again.
Our paper improves upon the best approaches for doing approximate planning in large problems in two ways :
It obtains tighter confidence intervals on the quality of a policy by incorporating a bound on the probability of reaching additional (not-yet-visited) states. These tighter intervals mean that fewer simulations are needed.
It introduces a more strategic method for choosing which state would be best to sample next by maintaining a discounted occupancy measure on all known states.
Our work is based on the idea of being able to restart planning from a fixed start state at any time. This idea was originally put forward by Fiechter in 1994 . However, many important innovations have been made in PAC-MDP community which we apply to our problem.
A fundamental feature of many PAC-MDP algorithms is optimism under uncertainty. That is, if there are some states we’ve never encountered or evaluated, then we assume they are high value states. This encourages the algorithm to try to reach unknown states and find out their true value. If such a state does indeed have high value then we benefit directly; if it’s a bad state, then we learn quickly and become less likely to visit the state again. For optimism under uncertainty to work, we need to have an estimate of how likely we are to encounter any particular state over time.
One way to do this is with a confidence interval on the probability for reaching a state. The confidence interval can be computed based on how many times we’ve visited that state in previous simulations. The confidence intervals used in previous algorithms are quite loose. Their width typically depends on the square root of the number of states in the the state space. In spatial ecosystem management problems the state space is exponentially large, so this leads to very wide intervals. However, another property of real-world problems can help us. Typically, when we apply an action in a state, the set of possible resulting states is small. This means that only a small fraction of all the states will actually be reached over the planning horizon. So we can use the Good-Turing estimator of missing mass to put an upper bound on the total probability of all the states we have never visited. We integrate this upper bound into the existing confidence bounds and get a tighter one that better represents when to stop exploring.
Since the key performance cost that we are trying to minimize is the cost of invoking the simulator, the key is to invoke the simulator on the most interesting state at each time step. The second advance in our paper presents a new way to define “most interesting state” by using an upper bound on the occupancy measure. The occupancy measure of a state is essentially an estimate of how important a state will be given how likely we are to visit it and how far it is from the starting state. More precisely, it is the discounted probability that the optimal policy will occupy the state summed over all time, starting from a fixed starting state. We can compute and update this occupancy measure by dynamic programming during planning. Our key observation is that we can estimate the impact of choosing to explore some state K with action A over another in an efficient way. We compute just the impact locally on the confidence interval for the value Kweighted by its occupancy measure. This lets the algorithm focus exploration on the most promising state-action pairs.
We ran our algorithm on four different MDPs including a form of the invasive species river network described above and compared the results to the optimal solutions and results from some other algorithms. We found that our approach requires many fewer samples than previous methods to achieve similar performance. The algorithm does this while maintaining standard PAC bounds on optimality.
If you want to know more you can read our paper here. To take a look at the invasive species river network problem yourself there is a more detailed problem definition and downloadable code from this year’s RL planning competition which included our problem as one of the test domains.
A Markov Decision Process(MDP)[1,2] is a standard mathematical formulation for decision making problems containing states describing the world, actions that can be taken in each state, rewards that represent the utility obtained for taking an action in a given state and dynamics which define a conditional probability of transitioning from one state to another given a particular action. The solution to an MDP is a policy that tells for each state what action to take in that state in order to optimize the long-term cumulative reward. Given an MDP we can define the value of a policy as the expected reward obtained by following the policy over an infinite planning horizon discounted so that states farther in the future have less impact on the value.
There is a wide literature on solving MDPs exactly. The computational cost of these methods scales as the product of the number of actions times the square of the number of states. One community of approximate method that are well studied are the Probably Approximately Correct MDP (PAC-MDP) methods[3,4]. These methods take the idea of PAC estimators from statistics and apply them to estimating the value function of a policy. Essentially, an algorithm for learning a policy is said to be PAC-MDP if we can show that there is at least a probability (1-δ) chance that the value of the policy is within ε of the value of the optimal policy. Further, the algorithm must be efficient: It must halt and return its policy within an amount of time that grows only polynomially in the sizes of the input variables.
Bellman, R. 1957. Dynamic Programming. New Jersey: Princeton University Press.
Puterman, M. 1994. Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley Series in Proba- bility and Mathematical Statistics.Wiley.
Fiechter, C.-N. 1994. Efficient Reinforcement Learning. In Proceedings of the Seventh Annual ACM Conference on Computational Learning Theory, 88–97. ACM Press.
Strehl, A., and Littman, M. 2008. An Analysis of Model- Based Interval Estimation for Markov Decision Processes. Journal of Computer and System Sciences 74(8):1309–1331.
Good, I. J. 1953. The Population Frequencies of Species and the Estimation of Population Parameters. Biometrika 40(3):237–264.