Yes, that’s true — I’ll update the article to reflect that point. It also serves to further highlight a benefit of Thompson Sampling; TS can help an agent learn much faster. Also, the use case that is the subject of the article — variant testing in online settings — tends to be time bounded, meaning getting stuck is still a risk.
Yes, that is correct. I chose (20, 20) because I knew these were no where near the true parameters — therefore forcing the model to learn from an initially distant state. Equally, I could have used a random number generator to select the numbers.
I wouldn’t get to hung up on this point though, random number initialization of…
Hi Alex- thanks for taking the time to read. You make excellent points. This is very incomplete and I agree that if you want to get into ML you need to understand bias/var tradeoff amongst a host of other things. I would also highly recommend ISL to new students of ML. This post is really a intro to what I consider the “bear essentials”.