I versus Algorithms

It started with a misadventure 2 years back, when driving back from dinner to our hotel in Ooty (hill station in Southern India), Google Maps led me to a dead-end in one of the narrow roads in the hill town.

It was drizzling and dark. There was not enough road space to take a U-turn, so my wife got off the car and called for some help. I had to reverse my car downhill for about 20 minutes down 50 meters of the slope, with the valley on my left. With luck, patience and some generous assistance from the local resident, my car got back on the road; with offline guidance, we reached our hotel.

The Maps experience left me intrigued; I tried to think of it more as a technology problem or, algorithms that power it, than with Google.

Algorithms all around us

Algorithms in simple terms are problem solving tools that run from aerospace, medical science, recommendations on most content to commerce sites, best candidates for a job fit to the Maps et al, with a predictive capability.

They are deployed to solve problems that have high mathematical complexity (e.g new drug research, monsoon or, weather forecasts), generate new insights from data sets (e.g consumer behaviour from purchase/usage pattern), as well as execute decisions unassisted in certain situations (e.g trades on a securities exchange linked to price movement/econometric data et al, or guiding a vehicle through a traffic).

The starting point of algorithms is not data, rather the problem that we seek to solve.

Be it the expected travel time from point A to B, or monsoon forecast the predictive capability is as good as the data, the modelling done, time period involved and the iterative process of correction (actual vs predicted) and elimination (variables with marginal contribution to predicted value).

So what went wrong?

As I looked for answers to my Ooty incident, I tried finding ingredients that go into Google Maps. While the company may be deploying significant computing power (AI, if you insist), taking much wider data inputs, fundamentally it uses:

A. Historical traffic data on time taken to travel on those routes at specific times of the day, and

B. Real-time traffic data from several smartphone devices en-route that have location access activated.

When we switch on the services it maps the geographical location, and uses historical with real-time data to provide travel time estimates.

Back in Ooty, when I switched on the services at about 10pm, there weren’t many vehicles on the road and not being a tourist-heavy part of the year, most vehicles were local and may not be using the Maps services. The real-time component was weak in its predictive input.

Secondly, the terrain and its several short roads may not have been well mapped, thus picking one road to a dead-end.

In short, the algorithm did not have enough data, or historic/geographic information to provide the service.

What the algorithm may have also mimicked from its developers or, humans, the simple inability to say – I don’t know!

Payback time

As I drove back home from office on the same route (around the same time) that I have been taking for last few months, the commute time for the 13kms unassisted distance was a punishing 64 to 75mins.

Looking for a wiser hand, I switched on Google Maps; with the benefit of congestion ahead alerts and alternative route recommendations, my commute time fell to 55 minutes!

While taking the guided road last week, I kept thinking of alternative routes, but never tried as I trusted being in good hands. Plus, my research said Bangalore as a city is well mapped with great historical traffic data, and the several on-demand cabs that keep their location services activated leading to high real-time inputs, both providing high quality predictive output.

Then came the tipping point.

While taking the guided route yesterday, I took a switch to an alternative route which I had known and I instinctively thought presented lesser traffic.

I was home in 43 minutes, a 22% more efficient path than a smart algorithm!

So, what may have led to that route not getting recommended or directed?

My guess: the high volume of vehicles with location-services enabled on the route provided more real-time inputs, leading the algorithm to choose those routes, while insufficient historical data or, traffic on alternative route reduced its predictive capability. Hence, the algorithm may be chosing predictive capability than fastest time between two points.

The Future

Back to my 3 options:

-          Go unguided on the same route

-          100% dependence on a technology

-          Use technology plus own instinct (prior knowledge)

We get exposed to new experiences in life, visit new places and in these situations we can either choose to stay unguided, or take some assistance.

However, we need to believe that tools we develop be it automobile or, AI are born out of the confines of our intelligence. Yes, they can help us to increase our intelligence, enable more for us and our societies, and take us into those parts of space to molecular research that we can’t reach, but human mind is capable of exponential change.

The best modelled programs/algorithms can provide solutions to crop research, healthcare, cancer cure and several other ambitious projects. We need a lot of them to advance our knowledge and that is the bedrock of human curiosity and existence. 

The hybrid road of human mind/intelligence with smarter technology is the collaborative future forward.

As we deploy more and smarter algorithms/AI, it is equally important to immerse technologies and learning tools to the deepest sections of our society, so that we don’t fear losing  low-skill jobs to technology/automation, but ride on them to generate new businesses, creative professions and unlock greater potential.