Consumer products contribute to more than 75% of global greenhouse gas (GHG) emissions, primarily through indirect contributions from the supply chain. Measurement of GHG emissions associated with products is a crucial step toward quantifying the impact of GHG emission abatement actions. Life cycle assessment (LCA), the scientific discipline for measuring GHG emissions, estimates the environmental impact associated with each stage of a product from raw material extraction to its disposal. Scaling LCA to millions of products is challenging as it requires extensive manual analysis by domain experts. To avoid repetitive analysis, environmental impact factors (EIFs) of common materials and products are published for use by LCA experts. However, finding appropriate EIFs for even a single product under study can require hundreds of hours of manual work, especially for complex products. We present Flamingo, an algorithm that leverages natural language machine learning (ML) models to automatically identify an appropriate EIF given a text description. A key challenge in automation is that EIF databases are incomplete. Flamingo uses industry sector classification as an intermediate layer to identify when there are no good matches in the database. On a dataset of 664 products, our method achieves an EIF matching precision of 75%.
Rule-based control (RBC) is widely adopted in buildings due to its stability and robustness. It resembles a behavior cloning methodology refined by human experts; however, it is incapable of adapting to distribution drifts. Reinforcement learning (RL) can adapt to changes but needs to learn from scratch in the online setting. On the other hand, the learning ability is limited in offline settings due to extrapolation errors caused by selecting out-of-distribution actions. In this paper, we explore how to incorporate RL with a rule-based control policy to combine their strengths to continuously learn a scalable and robust policy in both online and offline settings. We start with representative online and offline RL methods, TD3 and TD3+BC, respectively. Then, we develop a dynamically weighted actor loss function to selectively choose which policy for RL models to learn from at each training iteration. With extensive experiments across various weather conditions in both deterministic and stochastic scenarios, we demonstrate that our algorithm, rule-based incorporated control regularization (RUBICON), outperforms state-of-the-art methods in offline settings by and improves the baseline method by in online settings with respect to a reward consisting of thermal comfort and energy consumption in building-RL environments.
Products contribute to carbon emissions in each phase of their life cycle, from manufacturing to disposal. Estimating the embodied carbon in products is a key step towards understanding their impact, and undertaking mitigation actions. Precise carbon attribution is challenging at scale, requiring both domain expertise and granular supply chain data. As a first-order approximation, standard reports use Economic Input-Output based Life Cycle Assessment (EIO-LCA) which estimates carbon emissions per dollar at an industry sector level using transactions between different parts of the economy. EIO-LCA models map products to an industry sector, and uses the corresponding carbon per dollar estimates to calculate the embodied carbon footprint of a product. An LCA expert needs to map each product to one of upwards of 1000 potential industry sectors. To reduce the annotation burden, the standard practice is to group products by categories, and map categories to their corresponding industry sector. We present CaML, an algorithm to automate EIO-LCA using semantic text similarity matching by leveraging the text descriptions of the product and the industry sector. CaML uses a pre-trained sentence transformer model to rank the top-5 matches, and asks a human to check if any of them are a good match. We annotated 40K products with non-experts. Our results reveal that pre-defined product categories are heterogeneous with respect to EIO-LCA industry sectors, and lead to a large mean absolute percentage error (MAPE) of 51% in kgCO2e/$. CaML outperforms the previous manually intensive method, yielding a MAPE of 22% with no domain labels (zero-shot). We compared annotations of a small sample of 210 products with LCA experts, and find that CaML accuracy is comparable to that of annotations by non-experts.
Existing approaches for autonomous control of pan-tilt-zoom (PTZ) cameras use multiple stages where object detection and localization are performed separately from the control of the PTZ mechanisms. These approaches require manual labels and suffer from performance bottlenecks due to error propagation across the multi-stage flow of information. The large size of object detection neural networks also makes prior solutions infeasible for real-time deployment in resource-constrained devices. We present an end-to-end deep reinforcement learning (RL) solution called Eagle1 to train a neural network policy that directly takes images as input to control the PTZ camera. Training reinforcement learning is cumbersome in the real world due to labeling effort, runtime environment stochasticity, and fragile experimental setups. We introduce a photo-realistic simulation framework for training and evaluation of PTZ camera control policies. Eagle achieves superior camera control performance by maintaining the object of interest close to the center of captured images at high resolution and has up to 17% more tracking duration than the state-of-the-art. Eagle policies are lightweight (90x fewer parameters than Yolo5s) and can run on embedded camera platforms such as Raspberry PI (33 FPS) and Jetson Nano (38 FPS), facilitating real-time PTZ tracking for resource-constrained environments. With domain randomization, Eagle policies trained in our simulator can be transferred directly to real-world scenarios2.