Machine learning lecture 1  (Page 10/17)

And so over time, Arthur Samuel had a checkers program that would actually learn to play checkers by learning what are the sort of board positions that tend to be associated with wins and what are the board positions that tend to be associated with losses. And way back around 1959, the amazing thing about this was that his program actually learned to play checkers much better than Arthur Samuel himself could.

So even today, there are some people that say, well, computers can't do anything that they're not explicitly programmed to. And Arthur Samuel's checkers program was maybe the first I think really convincing refutation of this claim. Namely, Arthur Samuel managed to write a checkers program that could play checkers much better than he personally could, and this is an instance of maybe computers learning to do things that they were not programmed explicitly to do.

Here's a more recent, a more modern, more formal definition of machine learning due to Tom Mitchell, who says that a well-posed learning problem is defined as follows: He says that a computer program is set to learn from an experience E with respect to some task T and some performance measure P if its performance on T as measured by P improves with experience E. Okay. So not only is it a definition, it even rhymes.

So, for example, in the case of checkers, the experience E that a program has would be the experience of playing lots of games of checkers against itself, say. The task T is the task of playing checkers, and the performance measure P will be something like the fraction of games it wins against a certain set of human opponents. And by this definition, we'll say that Arthur Samuel's checkers program has learned to play checkers, okay?

So as an overview of what we're going to do in this class, this class is sort of organized into four major sections. We're gonna talk about four major topics in this class, the first of which is supervised learning. So let me give you an example of that.

So suppose you collect a data set of housing prices. And one of the TAs, Dan Ramage, actually collected a data set for me last week to use in the example later. But suppose that you go to collect statistics about how much houses cost in a certain geographic area. And Dan, the TA, collected data from housing prices in Portland, Oregon. So what you can do is let's say plot the square footage of the house against the list price of the house, right, so you collect data on a bunch of houses. And let's say you get a data set like this with houses of different sizes that are listed for different amounts of money.

Now, let's say that I'm trying to sell a house in the same area as Portland, Oregon as where the data comes from. Let's say I have a house that's this size in square footage, and I want an algorithm to tell me about how much should I expect my house to sell for. So there are lots of ways to do this, and some of you may have seen elements of what I'm about to say before.

So one thing you could do is look at this data and maybe put a straight line to it. And then if this is my house, you may then look at the straight line and predict that my house is gonna go for about that much money, right? There are other decisions that we can make, which we'll talk about later, which is, well, what if I don't wanna put a straight line? Maybe I should put a quadratic function to it. Maybe that fits the data a little bit better. You notice if you do that, the price of my house goes up a bit, so that'd be nice.