The Evolution of Machines
Can you imagine life without machines today? Technological advances driven by Automation/Computing and Communication are rapidly transforming the way we live our daily lives.
In the early days of the industrial revolution, it was the motorized machines which targeted tasks requiring manual labor – such as a tractor to plow our fields or a crane to lift heavy objects. These machines still needed a human to direct the machines. The decision-making process rested with a human, who would pull levers or press buttons to accomplish the task. The next evolution in this direction led to programmable machines, where we were able to define instructions and the machines would execute those instructions to operate. But what would happen if the machine encountered a situation which was not programmed? Or if the instructions were just too complex to define? For example, how do we program a machine to identify a human face?
There is an identifiable pattern in these problems that humans leverage to make decisions when presented with fuzzy data or a situation varying from the predefined protocol. With machines, the challenge is how to translate this Human Intelligence and train them. For long, Artificial Intelligence has been the subject of research for application to selective fields like Bio-metrics, Robotics, Finance, Medical Science and Industrial Automation, etc.
More recently the advances in Mobile Technology have unleashed a plethora of Smart devices that help track tiny details around our daily-lives like fitness/food habits/shopping habits at the Micro-level and the ecosystem around us like traffic-conditions/weather/economy at the Macro-level.
Today, Artificial Intelligence is playing a key role in leveraging this data to train devices that help us plan our daily-lives, better. The impact is visible everywhere starting from our homes, classrooms, cars, hospitals, and offices, etc.
Artificial Intelligence and Machine Learning
Machine Learning is a term often used interchangeably with Artificial Intelligence. Other terms you may often come across when reading up about AI are Deep Learning and Neural Networks. Let us start by defining the different terms and establish their relationships.
Artificial Intelligence: the science of making machines capable of taking human-like decisions.
Machine Learning: a technique used to train machines into imitating human-like decision-making process. So, basically, this is one of the processes we have available to make machines intelligent – though it is not the only process possible.
Neural Network and Deep Learning: Neural Networks are a class of Machine Learning algorithm. The idea behind these algorithms is to imitate the biological world where a network of simple unit emerges in a complex intelligence system. Deep Learning is a way of constructing neural networks so that there are multiple layers of neural units between input and output. This is significant as traditionally, machine learning required that the input data was cleaned up and presented to the machine learning units in a way that could be consumed easily (feature engineering). This involved teams of data scientists reviewing, interpreting and transforming large volumes of data for long hours. With the Deep Learning approach of using multiple neural layers, a large part of the cleanup is now done by layers of neural network themselves – thereby making it more practical to apply machine learning models to real-world problems.
So, to reiterate, Artificial Intelligence (AI) allows machines to make decisions as humans would. Machine Learning (ML) is a technique we deploy to build this intelligence in the machines. The process followed by ML is inspired by how a human child learns about the real world. A child learns when parents point to objects and label them, say to teach a child about cars, parents pointed to different car pictures and after some examples, asked the child to identify cars on the road. The child starts pointing at objects he has learned to identify as cars. At times when the child is correct, the parents encourage him/her, while at other times when a child points to a bus or a truck and says it’s a car, parents correct the child.
This method establishes a process through which children themselves unravel the patterns of what constitutes a car and what does not. In this example, parents would have had difficulty defining what a car is to a child. Nevertheless, by presenting examples, we can train a child’s brain. In the similar vein, teaching machines “what a car is” by describing its features would have been impossible. Defining it as an object with four wheels, an engine, windows, and a chassis, etc., would have first involved – defining what a wheel is, how to identify engine etc. But even after that description, there will still be confusion on what differentiates a bus from a car and a machine may miss out on correctly identifying an SUV (also known as Station-Wagon) as a car.
In its essence, reliable ML solutions leverage statistical methods, like regression, to train a model by presenting large volumes of data and crunching the data through fast computations. In the past, enabling individual computers to run sophisticated machine learning models was a costly affair. However, with Cloud platforms offering on-demand computational ability, it has now become economically feasible to teach machines and re-train them as needed.
Additionally, over the recent years, the amount of data being streamed and stored has grown exponentially. Smart devices fitted with sensors or GPS are connected to Clouds and keep pushing the stream of their events to the server. For instance, at MoveInSync, the driver device and the fixed device fitted into the cabs keep sending the cab’s location every few seconds or after traveling every few meters. With over 27,000 cabs being powered by MoveInSync devices, we are collecting GBs of data every day.
AI in Transportation and MoveInSync
AI has equipped the transportation industry with Control functions and other powerful tools like pattern recognition, rule-based decision systems, self-learning processes, etc. The Aircraft “Auto Pilot” is an AI application in use for few decades by most commercial airlines.
In addition to addressing on-route challenges by enabling “Autonomous vehicles” and “Traffic Management Systems” to make travel safer and time-efficient, AI is also being put to use to optimize resource usage.
Across the world especially in populous countries like India, several million employees are transported from home-to-work and vice versa. An Employee Transport Service Provider’s key offering is that it will bring all commuting-employees to office on-time and ensure employees do not have to wait when leaving office. To fulfill this promise, transport teams typically engage with cab vendors to supply vehicles. Clearly, there is a demand and supply balance to be maintained while ensuring other parameters like on-schedule-travel, safety and cost are not compromised. MoveInSync Employee Transport Management solution has been very effective at collectively managing these parameters.
Variation in demand by the employees along with the vast locational/temporal separations makes it challenging to ensure that cabs are available at every desired location only when needed. To counter this, transport teams build a buffer of cabs. Many times, these cabs spend long hours doing no business on some days, while on other days, even this buffer may fall short.
For instance, say for a 09:00 am Login shift at an office in Bangalore, typically 100 employees place schedule requests. However, these employees may be concentrated unequally across Bangalore city. For instance, 11 employees board from the Electronics City zone, so the transport team would arrange for an 8-seater vehicle like TATA-Sumo and a 4-seater vehicle like Indica to meet this demand. On days when the number of employees falls to 7, a Sumo alone would suffice leaving the Indica idle. On the other hand, on days when 14 employees schedule requests, 2 Sumos are required to fulfill the demand. The Transport team could capture employee schedules well in advance, but it has been observed that this typically leads to higher No-Shows by employees – thereby again leading to wastage. As a result, companies try to capture employee travel-requests as late as possible. This is with the aim of improving employee satisfaction and reducing wastage, thereby reducing the time available for the transport team to plan the cabs according to actual demand variation.
MoveInSync is leveraging AI to solve this problem by predicting the hourly demand in advance from any part of the city. Once the available cap supply is mapped to the predicted demand, any possible cab shortage can be identified early, and cabs arranged in time. The set of problems which require demand predictions are labeled Time Series Forecasting. In such problems, what happens today has a strong correlation with what’s going to happen in future. A typical time series has multiple periodic cycles (such as weekly cycle, monthly cycle or seasonal cycles) and white noise (normally distributed). By coupling the internal demand behavior with external factors like weather forecasts (especially, rain), local events, holiday calendars for the city (including long weekends) road-closures, public transport work schedules, etc. MoveInSync predicts the area wise demand. This has played a key role in enabling transport managers to utilize cabs to their maximum and ensure operational efficiency.
The MoveInSync Edge
Headquartered in Bangalore, MoveInSync today services 75+ clients across 20 cities through a fleet of more than 27,000 cabs running daily operations. In our endeavor to serve our customers better, MoveInSync Research teams have been creatively leveraging the historical operations data to constantly enhance the Forecasting model, thereby driving 2 of our core values – “Passion for Customers” and “Perseverance for Results”.
Vikas Sethi, Independent Consultant
Videh Ranjan, Director, Product Management