-1
$\begingroup$

There's many types of Machine Learning methods, but one sticks in my mind, and I don't understand why for ex. Q Learning was made instead of just using the following method:

Method: Say we give the robot some Triggers, we will implant a pre-installed image of a girl, taste of fries, sound of birds, and acceleration. When it Guesses Actions - that get a trigger, it saves actions just done and links them to senses just saved and labels the senses&actions as + or - based on if matched a + or - trigger and gives it a rating of how much it matched, then when it senses it again later, it matches memory and links to the actions and they are initiated, it walks to the girl/sound of birds chirping.

Seeing that, why is Q Learning so complex the way it works?

$\endgroup$
3
  • 2
    $\begingroup$ You sound very naive. Try implementing such a system to see what goes wrong. $\endgroup$ Commented Jan 7, 2016 at 23:03
  • $\begingroup$ By image matching I mean scaling/rotating/tranlating/ect and the more that matches the more it saves guessed actions. By Q Learning I meant what else besides finding pixel patterns could it be doing? $\endgroup$ Commented Jan 7, 2016 at 23:39
  • $\begingroup$ Again, you must explain what will "not work". My method isn't perpetual motion, it does-work. $\endgroup$ Commented Jan 7, 2016 at 23:43

1 Answer 1

2
$\begingroup$

What you are describing is similar to what virus detection software does. It is given a list of signatures and then searches for them in its input stream. In contrast, machine learning works in a setting in which you don't expect identical inputs to typically appear. For example, consider the task of recognizing digits. Here are some examples of the digit 2, from a paper of Mayraz and Hinton:

The digit 2

None of these is identical to the others. Suppose that you only train your learner with the upper half. When encountering a digit from the bottom half, how would you recognize it as 2? You need to generalize your training examples. This is what machine learning is about.

$\endgroup$
10
  • $\begingroup$ Mhm, I think you better re-analyze my whole opening post again, it isn't waiting for exact triggers to match, if it's CLOSE then it will save actions and the rating they are given will be lesser. Now, seeing this all, it works doesn't it? $\endgroup$ Commented Jan 7, 2016 at 23:01
  • 1
    $\begingroup$ This approach is known as "nearest neighbor" (q.v.), and is outperformed (in many cases) by algorithms such as SVM and neural networks. This is an empirical fact. $\endgroup$ Commented Jan 7, 2016 at 23:03
  • 4
    $\begingroup$ It will have poor performance in many cases, for example digit recognition. This is an empirical fact. Stating that it works doesn't make it so. At this point it is pointless to advance the discussion any further before you actually test your ideas and see if they have empirical merit compared to other approaches. $\endgroup$ Commented Jan 7, 2016 at 23:08
  • 1
    $\begingroup$ I mean digit recognition. For example, given a photo of an envelope with an address written on it, determine the zip code. This involves recognizing the individual digits making up the zip code. $\endgroup$ Commented Jan 7, 2016 at 23:11
  • 1
    $\begingroup$ @FriendlyPerson44 what Y/Filmus is saying is that of course all algorithms will try to match an input to a pattern and do an action. But how exactly this matching is done can be critical. Nearest-neighbour is one simple matching algorithm, but it cannot account for more complex inter-relations which other algorithms (like neural networks) can account for. In this sense they produce different results and different abilities of generalisation $\endgroup$
    – Nikos M.
    Commented Jan 8, 2016 at 10:21

Not the answer you're looking for? Browse other questions tagged or ask your own question.