5

I'm trying to create a tool that can estimate how many days a project will take to complete, given dozens or hundreds of factors and parameters.

I don't have much experience with Artificial Intelligence / Neural Networks / Machine Learning but from my understandings it is good for pattern recognising. If I make an AI program and train it by feeding it with information on projects we have already completed (factors and number of hours took), it should be able to learn to estimate future projects.

Does anyone know if this is a realistic approach?

Pupper
  • 2,315
  • 2
  • 22
  • 29
  • 5
    In order for this to be a realistic approach you will need a lot of data to train the algorithms and that training data will have to have all of the same parameters plus the outcomes (i.e. how long it took to complete the project). If you have the data, then you should be able to create predictive models. – Kiril Dec 13 '11 at 03:07
  • 2
    For most machine learning techniques, the number of training samples you need grows very fast with the number of input features (some claim: exponentially!), so to train a NN with "hundreds of factors" you might need millions of training samples. Do you have those? – Niki Dec 16 '11 at 11:11

1 Answers1

2

Honestly, I think you will be better off building an expert ANN for how long individual tasks may take. There will be more data, more frequently; which will allow for greater specialization by the system.

There are just way to many factors in the overall project for it to be able to accurately predicted without an enormous data set.

If the expert system proves successful, you could couple a governing ANN over the top, which takes the summed project time from all tasks (as predicted by the expert ANN) and some other project factors (such as total employees, time of the year, budget, etc), and trained on this data from previous projects; and predict from that.

Too much information can make for very noisy data, and non-sensible output in edge cases, especially in this case, where they may be a lack of abundant training data.

hiddensunset4
  • 5,825
  • 3
  • 39
  • 61