The carbon footprint of Google’s eco-friendly route prediction
Sustainable AI for climate change
Climate change is one of the largest battles we are currently involved in. AI for climate change is one of the most happening research domains where scientists propose various AI solutions to tackle a range of factors that affect sustainability. One of the dominant issues is vehicle emissions, given the steep increase in the number of automobiles used on a global scale. In order to tackle this, Jen Fitzpatrick with her Google Maps team recently proposed a new feature that could suggest an eco-friendly route to reduce vehicle emissions.
But we tend to overlook another dominant source of carbon dioxide emissions -training large AI models. How is this tradeoff handled by Google?
Let’s first understand how AI models can help predict vehicle emissions, analyze the ML carbon emission issue and then finally ideate on how we could help make the eco-route feature more sustainable than it actually might be.
How to predict vehicle emissions using AI?
It is possible to train machine learning models to predict carbon dioxide emission based on a lot of factors related to vehicle dynamics and the road link. The fastest highway route is not always eco-friendly. At a given moment, the interplay between these factors is used to suggest paths while also satisfying the travel time budget.
On analysis, it was observed that the average speed of the vehicle has the largest impact on the emission rate.
The highly transient and dynamic nature of the data makes the problem statement more complex.
The ML models predicting emission could be broadly classified into microscopic and mesoscopic. The microscopic models work based on driving characteristics in the context of each point of time, like instantaneous speed and instantaneous acceleration. On the other hand, the mesoscopic models use driving characteristics in the context of a period of time, such as average speed and proportion of acceleration/deceleration.
Keeping track of momentary changes in data and predicting at every instance is very complicated. Hence, mesoscopic models are widely adopted.
Recent models use Vehicle Specific Power (VSP) as the primary factor affecting the emission rate. VSP estimates the power demand on the engine during driving. It is calculated using second-by-second speed values in a driving schedule, along with information about the type of vehicle being operated. The distribution of VSP values calculated is then used by these models. There are also link-based emission models that work on various transient vehicle behavior data and use link cost function with shortest path planning algorithms. The factors based on which these eco-routing methods work could be broadly classified into 5 categories:
- Link attributes — the kind of road (road level and properties). Eg: expressways, urban freeways, national highways, etc.
- Link’s length — This data could be obtained using the geographic database.
- Width of the link — Covers the quality and level of the road.
- Fee/Toll States — Different toll states indicate different states of maintenance which could impact the driving pattern. Eg: Toll road, unknown, free, etc.
- Average Speed of link — current traffic condition which is the dominant factor of driving pattern.
Various decision tree algorithms are put to use to acquire the relationship between driving patterns and link attributes. A typical link-based emission modeling involves:
- Obtaining driving patterns by analyzing and clustering with large amounts of data
- Constructing a link-based classifier that determines the kind of driving pattern (driving speed, acceleration, engine stress, engine rpm, etc.) that a vehicle on a specific link behaves in.
- Calculating the link-based emissions by applying together emission information with the corresponding driving patterns.
Does the Google Maps new feature of eco-route suggestion actually decrease the overall carbon dioxide emission?
Concerning vehicle emissions, the Maps algorithm employs the methods discussed in the previous section and helps reduce the emission rate. But it is imperative to also consider the carbon dioxide emissions by these large and complex machine learning models during the training and tuning process.
Training a single AI model can emit as much carbon as five cars in their lifetime and it can contribute to around 3 percent of total carbon dioxide globally. In fact, environmentalists consider AI to be similar to oil industries where data is the oil. The computational and environmental cost of training the models is directly proportional to their sizes and complexity. The emission rate explodes when additional tuning steps are used to improve the final accuracy. Now, imagine the amount of emission by these models that require multiple levels of tuning and testing to handle dynamic data. Given this, does the prediction model helps to reduce the overall emissions or merely compensate for the amount it released during its development? Think.
Currently, Google has shifted all its operations to carbon-free where the non-carbon intelligence system executes tasks requiring high computations when there is high solar and wind energy available without affecting the working of the various applications. They have also designed their own processors (TPUs) which could be used in the cloud to train models where sustainable energy sources are utilized. This is a great initiative that is perfecting with time. Speaking with respect to the latest eco-route prediction feature, are there other possible ways that could be considered to help make it more sustainable?
What could be done to make vehicle emission models more sustainable?
The search for better accuracy in AI models would continue to exponentially increase with the demand for huge volumes of data to be processed along with long training sessions. This would lead to the need for more energy consumption and significant carbon emission. A simple way to reduce this is to shift the training sessions to the cloud that is hosted near to the locations where there is a more significant consumption of renewable resources. Transparency in developing AI models where the energy consumption and model performance metrics are documented could be helpful as a base reference to choose suitable architectures given a problem statement, hence reducing the trial and error basis of training and testing various combinations of model layers and hyperparameters.
The difference in electricity consumption between tuned and untuned algorithms is significantly high. Assuming Google uses an efficient AI algorithm for the eco-route suggestion, what more could be done to achieve the ultimate carbon reduction goal? My two cents-
- Owing to the dynamic nature of data, can the eco-route suggested in itself be made dynamic? By predicting the updated eco-friendly route using the instantaneous driving and link information at frequent intervals, a combination of microscopic and mesoscopic methods could be used. This could be useful in case of accidents and similar unexpected events that could potentially clog any traffic congestion-free road link. Not to forget the tradeoff between the accuracy and model emission for the number of predictions! Complex enough?
- The traffic rules and roadway links change with different geographic locations and weather patterns. Is it more effective for a single model to accommodate these differences or multiple models that could be used for clusters of regions that show highly similar patterns with respect to these factors?
- Given source and destination points, do we really need to predict the eco-routes from the start to the end? Similar to pruning out the noisy data, subgraphs of the map that either has no alternate routes to suggest or are highly likely less congested in general could be pruned out. This significantly reduces the size of data for which the model makes predictions.
- Based on the trip length, we could classify the journeys as long and short trips. For the short trips class, do we really need to predict or just proceed with the shortest path? For the long trips class, is it possible to get additional information from the user if these trips are for leisure road trips? If yes, we could override the emission-time tradeoff and suggest the best eco-friendly route.
Sustainable AI and the future
Companies are realizing the environmental impacts of AI and working on the tradeoff between AI to reduce carbon footprints in industries and AI in itself being accountable for a considerable amount of the same. For instance, Microsoft has launched AI for Earth that empowers organizations and individuals working to solve environmental challenges. Google aims to run the business on carbon-free energy everywhere, at all times. AI for Good Global Summit: Home is a digital platform where scientists across the world could connect, build and work together on identifying practical AI solutions that could accelerate United Nations’ Sustainable Development Goals. They conduct frequent sessions that discuss various aspects of the role of AI in climate change.
It is encouraging to see how we have shifted our focus to a more sustainable tomorrow. What are some of the other ways in which we could make Artificial Intelligence more environment-friendly? Let us know!
References
- Quantifying carbon emissions of Machine Learning — https://arxiv.org/pdf/1910.09700.pdf
- Study on Eco-Route Planning Algorithm and Environmental Impact Assessment — https://www.tandfonline.com/doi/pdf/10.1080/15472450.2013.747822?casa_token=S1cOvcMHBKMAAAAA:J6AV7bNy_Sug3FEZQQAtDyE0CJqwd7vHn_SJGelx5lcnx28KjGCEaF8T_O78ZKvzwR0sWt0flK5-
