Discovering Effective Policies for Land-Use Planning with Neuroevolution

TL;DR

Addresses land-use planning to minimize $ELUC$ while limiting land-use change; uses Evolutionary Surrogate-assisted Prescription (ESP) to learn a surrogate of long-term emissions from historical context–action–outcome data and to evolve prescriptors via evolutionary search that produce Pareto fronts; trains on LUH2 and BLUE within Project Resilience and demonstrates predictive accuracy and policy performance; finds that a Global NeuralNet surrogate and evolved prescriptors outperform simple baselines in the critical middle-change region, and offers an interactive demo and RHEA-based human-expertise integration; this work provides a proof-of-concept decision-support tool for climate-conscious land-use planning with potential extensions to finer-scale data and uncertainty quantification.

Abstract

How areas of land are allocated for different uses, such as forests, urban areas, and agriculture, has a large effect on the terrestrial carbon balance, and therefore climate change. Based on available historical data on land-use changes and a simulation of the associated carbon emissions and removals, a surrogate model can be learned that makes it possible to evaluate the different options available to decision-makers efficiently. An evolutionary search process can then be used to discover effective land-use policies for specific locations. Such a system was built on the Project Resilience platform and evaluated with the Land-Use Harmonization dataset LUH2 and the bookkeeping model BLUE. It generates Pareto fronts that trade off carbon impact and amount of land-use change customized to different locations, thus providing a proof-of-concept tool that is potentially useful for land-use planning.

Figures (10)

• Figure 1: The Evolutionary Surrogate-assisted Prescription (ESP) method for decision optimization. A predictor is trained with historical data on how given actions in given contexts led to specific outcomes. It is then used as a surrogate in order to evolve prescriptors, i.e. neural networks that implement decision policies by prescribing actions for a given context, resulting in the best possible outcomes
• Figure 2: Visualizing the differences in model behavior. Predictions for ELUC (tC/ha) are created for the Global models using synthetic data created by changing 100% of land-use type A (row) to 100% of type B (column) in a 1% sample across the range of latitude, longitude, cell area, and year occurring in the test data. The results are averaged for each conversion from A to B. The models generally agree on the sign of ELUC, which in turn aligns with the sign generated by the BLUE model, suggesting that the results are reliable. The RF model is not able to extrapolate to large values, resulting in low predictions; LinReg and NeuralNet are similar but differ numerically, presumably due to the differences between linear and nonlinear predictions
• Figure 3: Evolution of prescriptors with the Global NeuralNet predictor. ($a$) The Pareto front moves towards the lower left corner over evolution, finding better implementations for the different tradeoffs of the ELUC and land-use change amount objectives. ($b$) Each prescriptor evaluated during evolution is shown as a dot, demonstrating a wide variety of solutions and tradeoffs. The final Pareto front (collected from all generations) is shown as blue dots. It constitutes a set of solutions from which the decision-maker can choose a preferred one
• Figure 4: The Pareto fronts of Evolved Prescriptors vs. heuristic baselines, with ELUC and land-use change evaluated on the Global test set. The Evolved Prescriptors achieved better solutions than the baselines in the middle-change region where the land-use changes matter the most, demonstrating that they can take advantage of nonlinear relationships in land use to discover useful, non-obvious solutions
• Figure 5: Comparing a selected Evolved Prescriptor with the Perfect Heuristic. ($a$) The average performance of the Evolved Prescriptor dominates that of the heuristic. ($b$) The averages are expanded into actual samples in the test set (subsampled for readability). The samples for the Even Heuristic are largely hidden under the samples for the Perfect Heuristic. The Evolved Prescriptor suggests many more large changes than the heuristic. ($c$) The differences in change percentage and ELUC between the Evolved Prescriptor and the Perfect Heuristic for each test sample, with color indicating which solution dominates. Surprisingly, this particular Evolved Prescriptor dominates the Perfect Heuristic only on a single individual sample. Thus, evolution discovered the insight that it is possible to do well globally by utilizing a few cases where large change is possible