A safe reopening of shopping malls with artificial intelligence & IoT support post Covid-19

Original article was published on Deep Learning on Medium

A safe reopening of shopping malls with artificial intelligence & IoT support post Covid-19

A representative shopping mall floor

While the world prepares to get back to the new normal post this current pandemic, one key task is reopening of the shopping malls. In India (and similar other countries), shopping malls have become the most coveted destination, be it for picking up staples or checking out the latest electronic gadget. But the pertinent question now is — when can we safely visit our nearest shopping mall again?

Let us imagine a safe experience of a patron (customer) in a shopping mall floor powered by next-generation deep learning and IoT technologies. For ease of visualization, please refer to the attached schematic of a typical shopping mall floor, oversimplified for this writeup.

Following is some key pointers around a patron journey and delivering a safe experience:

  • A patron arrives at the door of the shopping mall
  • Post preliminary checks, the security personnel hands over a box size device with headphones. Imagine this to be similar to the audio players given on rent to visitors in art galleries and museums.
  • The box contains a RASPI device with a microcontroller (Arduino), sensor modules, and accessories such as headphones, buzzer, microphone, LED display, and similar others.
  • Patron navigation through the floor is guided by audio-visual queues emitted from the device.
  • Each device contains a GPS sensor that transmits the movement of the patron (and hence changing relative distance) over a period of time to a central server.
  • The outlined ecosystem in the adjacent diagram may be conceived as a multi-agent reinforcement learning (MARL) model where the shopping floor is the environment, each patron is an agent, a reward is “safely” reaching the destination shop or other interim milestones.
  • A patron’s RASPI device learns through a reward mechanism (typical RL implementation) as they navigate through the floor. Some sample rewards (positive or negative) are: maintaining a safe distance with other patrons or accidentally touching a door handle in the washroom even though it may be sensor-based and meant to operate on its own, touching an ATM keypad which has not been cleaned post previous use (let us assume a person is designated to clean the keypad post each usage and mark as safe for the next customer).
  • In the Indian context, with the usage of the Arogya Setu app (https://en.wikipedia.org/wiki/Aarogya_Setu) being made mandatory, it should be possible to expose interfaces to detect a “risky” person in proximity.
  • The learning of one patron journey (model train) may be persisted within the device such that the next patron who uses the same device may benefit from it as much as possible.
  • Application of modern game theory, especially static games (where each player makes decisions simultaneously without the knowledge and strategies of other players) in cooperative, competitive or mixed-mode will make the learning efficient and effective where each player (patron) may benefit from each other’s traversal journey through the floors.
  • Patrons may be guided visually through a small LED screen attached to the device to walk through the best possible trajectory to reach their destination in the safest possible way.
  • On leaving the mall the patron would return the device back to the security to be sanitized and reused by the next patron

Deep learning models (MARL) do have a challenge of training time and efficiency when the number of agents becomes higher. However given the general awareness on social distancing there will not be too high a number of patrons on a single floor at a given point in time. Moreover, there is perhaps no need to model the entire floor for a large shopping mall. It can be logically divided into smaller segments or immediate vicinities around a given locus to build the learning just around the area of relevance.

On an ending note I would like to acknowledge the blog by Jesus Rodriguez at https://towardsdatascience.com/modern-game-theory-and-multi-agent-reinforcement-learning-systems-e8c936d6de42. This helped to understand how game theory concepts can be adopted into deep (multi-agent) reinforcement learning models.