“Self-driving cars are the natural extension of active safety and obviously, something we should do”
~ Elon Musk

    


Driving any vehicle safely requires a constant stream of quick accurate decisions from the driver, be it measuring the braking distance needed to stop at a point, gauging the distance from a specific object, or negotiating with traffic. In the absence of a driver in autonomous vehicles, the responsibility of making these critical decisions falls on the Artificial Intelligence (AI) controlling the vehicle.

Therefore, it is crucial for the decision-making system to be extremely accurate. A significant role of the decision-making system is calculating probabilities, for example the probability of another vehicle entering the lane occupied by the car, a pedestrian crossing the road, or any other sudden obstruction on the road.

Prediction by Models

In order to predict the behavior of the surrounding cars, the AI in an autonomous car uses prediction algorithms based on models. This is how they work-

  • Multiple sensors in the car (such as Radar and LIDAR) collect data regarding the environment around the car, which is compiled and sent to the AI to run various prediction algorithms
  • A localization system tells the AI exactly where the car is in relation to the cars around it.
  • Now in order to get a sense of what nearby traffic will do, a model is made for all the possible trajectories the nearby cars can take (turn right or left, change speed, change lane). These models are sent to a Multi Modal Estimation Algorithm which assigns probabilities to each trajectory model based on likelihood of occurrence.
  • A process model is created for all the possible behaviors and trajectories that the nearby traffic can take.
  • In order to estimate the likelihood and accuracy of the model, the state of the traffic is observed at time x and time x-1. Now, the model is run at time x-1 and the expected state at time x is observed.
  • The observed state of the traffic and estimated state are compared and thus we get a probability for the model.
  • Finally, the trajectory model with the highest probability is selected and the car follows that trajectory.

 

How sensitive are these models?

These models need to be precise and accurate, as their outcomes affect people’s lives, both inside and outside the car. Moreover, the data that is sourced from the sensors needs to be extremely reliable.

For example, the model estimates the probability that another car will enter the lane where the self-driving car is as 0.01, but the actual probability is 0.1. This seemingly small error has the potential to have fatal consequences.

Thus, there is a high degree of sensitivity in these models as the slightest change in the input data or estimations can lead to devastating results.

Embracing probabilities in autonomous cars.

Not a lot of autonomous car-makers have embraced probabilities into their AI, but instead have gone for other approaches. There are a host of reasons behind this:

  • Uncertainty

The use of probabilities will raise concerns in the minds of the people. People are willing to adopt self-driving cars only if they have a 0% chance of crashing. The thing is, even if the self-driving cars’ crash rates are lower than those of conventional cars, individual crash cases will be highlighted and used to argue that self-driving cars are unsafe. Only if these new-age cars are brought into the roads and tested will they ever have a chance of attaining a 0% crash rate.

  • Lack of probabilistic programming languages

Until recently, AI developers did not have access to a lot of programming languages which could imbibe probabilities.

To fulfil this need, Uber has developed a probabilistic programming language Pyro, which is based on the Python programming language. Pyro can be used for deep probabilistic modelling, which unifies deep learning and Bayesian modelling. It is an open-source language, meaning that anyone can modify or build upon it.

A Level 5 self-driving car is a car which requires no human intervention and can drive on all terrains in any weather conditions all year round. To achieve this milestone in automation, it is essential that probabilities and uncertainties are coded into the AI of the car. Doing so will also enable the AI to continuously keep improving on itself.


Sources -

https://www.aitrends.com/ai-insider/probabilistic-reasoning-ai-self-driving-cars/

https://towardsdatascience.com/prediction-in-autonomous-vehicle-all-you-need-to-know-d8811795fcdc

https://futureoflife.org/2017/03/13/self-driving-cars-probability/#:~:text=Autonomous%20systems%20such%20as%20self,data%20collectors%20and%20reasoning%20algorithms

https://pyro.ai/


~ Kshitij Srivastava

Comments

Post a Comment

Popular posts from this blog

When An Actuary Helped Cricket Survive!!

Image
  When An Actuary Helped Cricket Survive!! Call for help, Comes an Actuary!! South Africa needed 22 off 13 balls to beat England in Sydney on March 22,1992 when rain stopped play. Ten minutes later the players were back on... and South Africa needed 21 off one ball. Blame the lowest-scoring-over rain rules, which ruined a cracking contest and chances of a WC Final. South Africa Team was out of the competition, but they left with a question hanging amongst every cricket fan…. What if this happens again??? That day Mr. Christopher Martin Jenkins said on the radio… "Surely someone somewhere could come up with something better” and soon Mr. Duckworth realized that it was a mathematical problem that required a mathematical solution." And then came the revolution!! ……. Since then, Mr. Duckworth – An Actuary , and Mr. Tony Lewis, a renowned Statistician worked together for years, just to bring out an effective method of calculating target score in 1997, officially
Read more

Operational Risk

Image
Operational risk can be defined as "the risk of a change in value caused by the fact that actual losses, incurred for inadequate or failed internal processes, people, and systems, or from external events (including legal risk), differ from the expected losses.” In simpler words, it is a risk faced by a company due to failed internal processes such as those of a faulty system, unsuccessful procedures, employee errors, and legal hazards. Some other factors include technological glitches such as breaches of personal data resulting from cybersecurity attacks or risks tied with automation and artificial intelligence. These can also be caused due to physical catastrophes such as cyclones, earthquakes, or man-made events such as terrorism and vandalism. Operational risk includes all other risks except those of credit and market. These risks can be as minor as an affordable loss caused due to poor management of human error; for example, a poorly trained employee may lose a sales opportuni
Read more

Market Risk

Image
  RISK Our life is full of risks. Everyone in some way or other comes across a risk . We often hear  people around us saying “This investment is a risk” or “I took a risk by moving to a new country”. But what exactly is a risk? There are various ways to define it. However, the simplest explanation of risk is “the possibility of something bad happening.”  Risks come in all shapes and sizes. Financial risks are in the more prominent types of risks. People are hesitant in making investments with good returns as they are reluctant to bear the risk due to insufficient knowledge of risk management which makes it necessary to educate ourselves about these financial risks and learn to manage them for our financial independence. MARKET RISK One of the most crucial financial risks is “Market Risk”. As the name suggests, it is a risk that affects the entire financial market i.e. it affects entire industries as a whole. Basically, it is the risk of facing losses due to unfavorable changes in the m
Read more