Route Optimization: Graphcore IPUs Solve Traffic Puzzles

Artificial intelligence is proving to be a powerful and versatile tool. In financial services, AI can fight credit card fraud and detect market sentiment to help investors. Algorithms can generalize cyber threats to make computer networks more secure and, in the process, even create branded audio for developers of hands-free devices. But there are some issues where conventional AI methods can fail. And one of those use cases is route optimization.

Predicting travel times in congested cities is a notoriously complex task. Developers face a moving target where scenarios can change quickly. Flowing traffic can suddenly come to a halt – for example, if a delivery van temporarily blocks a road – with cascading effects. AI can digest historical data and identify patterns, but first responders, courier companies, commuters and other city dwellers need real-time guidance.

IPU boost with GNNs

This is where Graph Neural Networks (GNNs) – deployed on efficient Intelligence Processing Units (IPUs) – come into play. A graph, in this context, is a structured representation of data and can be described as a network of interconnected nodes. Here, the nodes are individual road links and the links between them (called links) represent the relationship between these features.

Graphs allow developers to assemble a detailed picture of the city by encoding various information at each node. And running this structure through a neural network aggregates the data in the context of neighboring nodes. The process takes a complex relationship, such as the interactions between traffic elements across a city, and prepares the information so that it can be understood and optimized.

Running these routines on dedicated IPUs – which have been designed to hold full machine learning models in their processors – opens the door to additional insights. IPUs have been shown to speed up computations compared to using graphics processing units (GPUs) and are particularly good at speeding up GNNs despite their complex graph structures. This speedup is especially valuable when the end goal is to optimize real-time journeys.

timely advice

Another reason for using IPUs on intelligent transportation systems is the ability to perform so-called Mixture of Experts (MoE) traffic model integration. Rather than relying on a single global view, different behaviors can be introduced into the pipeline, for example, peak hour conditions or weekend patterns. The benefit is improved prediction accuracy. But with GPUs, the computational load of supporting multiple models would slow down the system. This is not the case for UPIs.

“[The IPU] allows multiple instructions and multiple data to be processed across different tiles,” comments Chen-Khong Tham, a professor at the National University of Singapore, who has used MoE model integration to improve traffic forecasting. “It’s very useful when you have operations that aren’t seamless.”

Recognizing that traffic patterns change over time, Tham and his colleagues used a sandwich workflow that takes into account both spatial and temporal characteristics. In the final step, AI is deployed to predict future traffic speeds on different stretches of road to aid in route planning and trip optimization. Through the use of IPUs – developed by UK company Graphcore – the team observed up to 4X speedup (compared to GPUs) in processing data from over 39,000 sensors distributed across major metropolitan areas from the California State Highway.

Unleashing urban traffic through route optimization – made possible by graphical neural networks running on IPUs – could have significant economic benefits. Online delivery companies, an industry that has accelerated in recent years, could improve their fleet management capabilities. The models could also respond to new factors such as the emergence of clean air zones and other urban planning initiatives.

GNI Trends

Graphical data structures and neural networks prove to be a handy combination in a range of applications. Graphs are a useful way to add structure to AI models so they can operate efficiently and gather locally relevant functionality – for example, by replicating the success of image recognition algorithms. Compared to using AI alone, adding graphics can help model new systems, provide data efficiency, and speed up training.

Combining these advantages with the tailor-made fit of IPUs that were designed with massively parallel processing in mind, provides developers with a very powerful analytical combination. And that is why GNNs running on IPUs are applied in various industries, not just transportation and “smart city”. Other areas of application include molecular analysis, drug discovery, stock market forecasting, social network analysis, and recommendation systems for e-commerce, to name a few examples.

To learn how IPUs can improve your AI workflow and provide noticeable processing speedup, visit https://www.graphcore.ai.