Reinforcement Learning

What is RL?

Learn through

  • Experience
  • Trial & error

How does it work?

  • Take an action
  • Get rewarded/punished
  • Learn from it
  • Improve your actions

Similar to how a child learns to walk

Image source: strong.io

Why RL?

Other ML algorithms are useful to find patterns

  • Lots of data
  • Complete data
  • Good understanding of the environment
  • No sequential decision making

In real-world

  • Complex problem
  • Partial information
  • No understanding of the environment
  • Dynamic system
  • Sequential decision making

ML Examples

Image source: medium @rohitlal & datacamp.com

RL Application

Image source: ICML 2019

Retail Sector

User-centric Recommender System

  • News
  • Product
  • Service
  • ...

Image source: ICML 2019

Technology Sector

  • Architecture selection
  • Device placement
  • Data augmentation
  • Cluster scheduling
  • Image source: ICML 2019

Energy Sector

  • Data center cooling
  • Smart grid
Image source: ICML 2019

Finance Sector

  • Option pricing
  • Order book execution
Image source: ICML 2019

Healthcare Sector

  • Dynamic treatment strategies
  • Medical image reporting generation
Image source: ICML 2019

Transportation sector

  • Dynamic seat pricing
  • Rideshare order dispatching

Impact

Adds personal touch

  • Hyper-personalization
  • Personalized offer/message @ right time/place/price

Avoiding ad overload using RL

  • Avoid showing same ad too often
  • Asses reactions to messaging and determine ideal frequency

Many success stories

  • FirstCry
    • Increased sales 10 folds
    • New customer rate up 10%
  • Alibaba
    • Advertisement auction, 240% ROI

References

RL4RealLife

Reinforcement Learning in Marketing

Lyft Marketing Automation

How Reinforcement Learning Boosts Marketing Performance

Machine Learning