Can’t get a taxi? We can relate.
But that’s only half the reason ride-sharing services are such a success.
It turns out if you’re a taxi driver you have an even bigger problem: it’s hard to predict where, and when to find passengers, Shin Ishiguro, a data scientist with Japanese telecom NTT DOCOMO, explained Wednesday at the GPU Technology Conference.
While ride-sharing services as most of the world knows them aren’t permitted in Japan, they’re a big attraction for drivers since they can just check an app on their phone to know where to go to get their next fare.
And while dispatchers can send taxis to people who request a ride, they can’t tell taxi drivers where to lurk to find places were people are most likely to hail them.
To bridge that gap, Ishiguro and his team at NTT DoCoMo created an app that helps taxi drivers go where the passengers are — and it’s already putting more money in the pockets of hard-working drivers.
Rather than relying on apps to connect amateur drivers with passengers, NTT DoCoMo data scientists use anonymized data from cell phone signals to detect where people are congregating.
It then uses deep learning algorithms running on an NVIDIA DGX-1 supercomputer to map that data to information collected by taxi drivers about demand for their services.
To predict demand based on cell phone usage, NTT DoCoMo uses stacked denoising autoencoder algorithms — a relatively new deep learning technique that allows machines to tease patterns out of seemingly random input.
The company delivers information via an app about where passengers can be found to an app that taxi drivers can access from a touchscreen.
The results have been uncannily accurate. Rolled out February 15 to 1,350 taxis in Tokyo and another 1,150 taxis in Nagoya, Japan, NTT DoCoMo’s app has steered taxis toward passengers with an accuracy rate of 92.9 percent.
That accuracy means drivers in the program have pocketed 1,409 yen more, or $13.15, for every day of driving than taxi drivers who aren’t. Over days and weeks of driving, those yen add up.