Figure neural network makes Google maps more intelligent

For the field of public travel , The arrival time of vehicles is the main influencing factor , Estimated time of arrival （ETA） Accuracy has become a very practical research topic . In recent days, , British artificial intelligence company DeepMind Deep cooperation with Google Maps , Using graph neural network （Graph Neural Networks,GNN） etc. ML technology , It has greatly improved Berlin 、 Tokyo 、 Real time in big cities like Sydney ETA Accuracy rate .

Many people use Google Maps （Google Maps） Get accurate traffic forecasts and estimated arrival times （Estimated Time of Arrival,ETA）, This is a very important tool , Especially when you are going through a traffic jam or need to attend an important meeting on time . For car sharing service companies and other enterprises , They can be used Google Maps The platform obtains the pick-up time information and estimates the price based on the ride time . DeepMind And Google Maps The research team cooperates , Try advanced machine learning techniques such as graph neural network , Upgrade Berlin 、 Jakarta 、 Sao Paulo 、 Sydney 、 Real time in Tokyo and Washington, D.C ETA Accuracy rate , The highest rise 50%. The following figure shows the ETA Promotion rate ：

Google Maps How to predict ETA For calculation ETA,Google Maps The real-time traffic data of different road sections around the world are analyzed . These data are Google Maps It provides an accurate picture of the current traffic conditions , But it can't help the driver to predict the driving time 10 minute 、20 minute , still 50 minute . therefore , In order to accurately predict future traffic conditions ,Google Maps Use machine learning to combine real-time traffic conditions of global roads with historical traffic patterns . This process is very complicated , For many reasons . for example , There are morning and evening rush hours every day , But every day 、 The exact peak time of each month is very different . Road quality 、 The speed limit 、 Traffic accidents and other factors also increase the complexity of the traffic prediction model . DeepMind Team and Google Maps Cooperate to try to improve ETA Accuracy rate .Google Maps For more than 97% The itinerary has a precise ETA forecast ,DeepMind And Google Maps The goal of our collaboration is to minimize the remaining inaccuracies in our forecasts , For example, Taichung （Taichung） Of ETA The prediction accuracy has improved 50% many . In order to achieve this goal on a global scale ,DeepMind A general machine learning architecture is utilized —— Figure neural network （GNN）, Spatiotemporal reasoning is carried out by adding relationship learning bias to the model , Then the connectivity of the real world road network is modeled . The specific steps are as follows ： Divide the roads in the world into super sections （Supersegment） The team divided the road network into several adjacent sections 「 Super section 」, The super sections have great traffic flow . at present ,Google Maps The traffic forecasting system consists of the following components ：
1） Route Analyzer ： Have a number TB Traffic information , It can be used to build super sections ;
2） new type GNN Model ： Use multiple objective functions for optimization , Be able to predict the travel time of each super section .

Google Maps Diagram of the model structure for determining the optimal route and travel time . Traffic forecasting with a new machine learning architecture Create a machine learning system for estimating travel time using super links , The biggest challenge is architecture . How to represent the variable scale sample of the link with any accuracy , So as to ensure that a single model can predict successfully ？ DeepMind The team's initial proof of concept began with a straightforward approach , This method makes best use of the existing transportation system , Especially the existing road network segmentation and related real-time data pipeline. This means that the super section covers a set of sections , Each section has a specific length and corresponding speed characteristics . First , The team trained a fully connected neural network model for each super road section . The preliminary results are good , It shows that neural network has great potential in predicting travel time . however , Given the variable size of the Super Section , The team needs to train the neural network model separately for each super section . To achieve large-scale deployment , Millions of such models must be trained , This poses a huge challenge to the infrastructure . therefore , The team started working on models that could handle variable length sequences , For example, cyclic neural network （RNN）. however , towards RNN Adding structures from the road network is difficult . therefore , The researchers decided to use graph neural networks . When modeling traffic conditions , How vehicles cross the road network is the focus of this study , Graph neural network can model network dynamics and information transmission . The model proposed by the team regards the local road network as a graph , Each road section corresponds to a node , Connect the two sections （ node ） The sides of are either on the same road , Or through the intersection （ Crossing ） Connect . When executing a message passing algorithm in a graph neural network , The message it transmits and its influence on the states of edges and nodes are learned by neural network . Look at it this way , Super sections are road subgraphs randomly sampled according to traffic density . therefore , Using these sampled subgraphs can train a single model , And a single model can be deployed on a large scale .

Figure neural network through generalization 「 Similarity degree （proximity）」 Concept , The learning bias imposed by convolutional neural network and cyclic neural network is extended （learning bias）, And then the connection with arbitrary complexity , It can not only deal with the traffic situation in front of and behind the road , You can also deal with adjacent and intersecting roads . In graph neural network , Neighboring nodes pass messages to each other . While maintaining this structure , The researchers applied a local bias , Nodes will be easier to rely on neighboring nodes （ This requires only one messaging step ）. These mechanisms enable the graph neural network to make more efficient use of the connectivity structure of the road network . Experiments show that , Extending the scope of consideration to adjacent roads that do not belong to the main road can improve the prediction ability . for example , Consider the impact of traffic congestion on the main road . By crossing multiple intersections , The model can predict the delay at the turning 、 Delay caused by merging , And the travel time of stop and go traffic conditions . The generalization ability of graph neural network in combinatorial space makes the modeling technology of this research have powerful ability . The length and complexity of each super section may vary （ From a simple two-stage path to a longer path with hundreds of nodes ）, But they can all be processed using the same graph neural network model .

from ‍ Basic research to production level machine learning model In academic research , Production level machine learning systems have a huge challenge that is often overlooked , That is, the same model will have huge differences in multiple training runs . Although in many academic studies , Subtle differences in training quality can be simply used as poor Initialization is discarded , But the subtle inconsistencies of millions of users add up to have a huge impact . therefore , When the model is put into production , Figure the robustness of neural network to this change in training has become the top priority . Researchers found that , Fig. the neural network is particularly sensitive to changes in the training process , The reason for this instability is that there are great differences between the graph structures used in training . A single batch diagram can range from two node plots to 100 The larger picture above the node . However , After trial and error , Researchers have adopted a new reinforcement learning technique in a supervised setting , Solved the above problems . In the process of training machine learning system , The learning rate of the system determines its own understanding of new information 「 Plasticity 」.

Over time , Researchers often reduce the learning rate of models , This is because there is a trade-off between learning something new and forgetting the important features that have been learned , Just like the growth process of human beings from children to adults . therefore , After a pre-defined training phase , The researchers first used an exponential decay learning rate scheme to stabilize the parameters . Besides , The researchers also explored and analyzed the model integration technology that has been proved effective in previous studies , So as to observe whether the model difference in training operation can be reduced . Last , Researchers found that , The most successful solution is to use MetaGradient To dynamically adjust the learning rate during training , Thus, the system can effectively learn its own optimal learning rate plan . By automatically adjusting the learning rate during training , This model not only achieves higher quality than before , But also learned to automatically reduce the learning rate . Finally, a more stable result is achieved , The new architecture can be applied to production . Model generalization through custom loss function Although the ultimate goal of the modeling system is to reduce the error in travel estimation , But the researchers found that , Using multiple loss functions （ Appropriate weighting ） The linear combination of has greatly improved the generalization ability of the model . To be specific , The researchers use the regularization factor of the model weight 、 Global traversal time L_2 and L_1 Loss 、 And for each node in the graph Huber And negative log likelihood （negative-log likelihood, NLL） Loss , Set a multiple loss target . By combining these losses , Researchers can guide the model and avoid over fitting the training data set . Although there is no change in the quality measurement of the training process , But the improvements that occur in training are more directly translated into setting aside （held-out） In test sets and end-to-end experiments . at present , Researchers are exploring , Under the guidance of reducing the travel estimation error ,MetaGradient Whether the technique can also be used to change the composition of the multi-component loss function in the training process . This research is supported by previous successes in reinforcement learning MetaGradient Inspired by the , And early experiments also showed good results .