1

I have got a problem where I have to automate some task. Let's say switching on and off an appliance based on the user interaction with it on different days of the week.

Let's say we have a bulb B1 and a user U1. In beginning U1 will switch on and off the bulb B1 at times he wants. For example he may switch it on at 9:10 AM on Sunday and switch off at 11:15 AM on Sunday. Similarly if we divide 7-days into 5 time units we shall get 2,016 unit times. Our system has to learn user behavior against each 5-minutes unit and after 2 weeks of learning switch on or off the bulb as the user would have.

Do you think it can be accomplished with Re-enforcement based learning techniques? Any other machine-learning algorithm suggestion?

It is basically for a home automation with machine learning university project.

Rehan Ullah
  • 111
  • 3

1 Answers1

2

When reading your question, my first intuition was to try a Bayesian Inference approach using a uniform distribution over the unit times as prior and user behaviour as evidence. A good visual explanation can be found here. Also, you can have a look at this medium post.

Sorin Dragan
  • 510
  • 4
  • 9
  • If I am getting you what you are trying to say is to use Bayesian Approach. Assigning probability of 1/2016 to each unit? Then treat each user event as evidence and updating along the way? – Rehan Ullah Nov 19 '20 at 12:35
  • Could you kindly give a bit more explanation? – Rehan Ullah Nov 19 '20 at 12:38
  • Yes, but keep in mind this was just a first thought. If your prior is uniform, your posterior will highly be influenced by your evidence. A simple way to model this is to count each time the user switched on (for example) the light in each time unit along the learning period of 2 weeks. You divide each count by the total number of observations to generate a probability mass function. Sampling from it should somehow mimic the user's behaviour. If you want to start with a different prior, you can have a look at the links in the initial response. – Sorin Dragan Nov 19 '20 at 13:11