A general idea of where machine learning is applied is readily available. But in which case where even when a rich training set is provided, the problem can't be solved through a machine learning algorithm? Where they shouldn't be used and why?
The best example would be predicting stock prices as they depend only on the current demand and availability of stock. Even when abundant historical stock rates are available to the public, machine learning methods cannot be applied on them for future price prediction.
I would like to know more of such practical examples.
Dear friend
Engineering problems like the above which is stochastic in nature, The machine learning is able to predict but the accuracy may be upto certain extent.
If any deterministic models are there, then ML will do the best than any other traditional models.
I hope this helps you.
Best regards
Dr.Indrajit Mandal
Machine learning needs an input to be trained on, in order to predict an output. So a very important thing is that there is a correlation between the input and output in some way.
You cannot train an algorithm to predict the number of ships sailing through the Suez, based on the number of pidgeons on St. Mark's Square in Venice.
So to make a bold statement, ML could in theory, learn to predict anything, you just need a suitable algorithm, the correct training data and efficient preprocessing. However this is probably more of a philosophical statement, than a scientific, and I assume that your intention was not to spur a discussion on epistemology.
With rather complex examples, it gets very difficult to figure out what training data you need. For the example you've given, the training data might consist of the frequency of wars, the number of years of a republic president in the us, the weather conditions and harvests, the speed of technological innovation.. and tons of other parameters.
Most ML algorithms assume an underlying distribution that does not change or is slowly varying in nature which can then be approximated as a static distribution. Changing distributions such as adversarial settings falls into examples of scenarios in which the algorithms wil fail if the adversary has information to circumvent the ML algorithm used(such as intrusion detection systems).
- Badges
- Science topic