Machine Learning

Machine learning is a subset of artificial intelligence, machine learning algorithms make mathematical models based on data to make decisions and predictions. The adaptability of machine learning makes it useful for tasks when conventional algorithms would not be suitable.

There are many types of machine learning some include supervised learning, semi-supervised learning and unsupervised learning. Supervised learning is when the algorithm is fed data with the input and the desired output. Semi-supervised learning is where it is fed an incomplete set of data and may have missing outputs or inputs. Unsupervised learning is the hardest and gives the computer only inputs and looks for commonalities and data structures and compares them to new sets of data.

This type of A.I. is becoming increasingly more important and useful as many complex tasks and predictions need to be carried out. It used in hospitals, space, genomics and more to analyze large data sets. This technology is being developed and tested in different ways, but one way is games. Machine learning was used to beat the world champion at go and chess which are complicated turn-based strategy games, but the scenarios are calculable, and you can see the whole board. However recently it attained one of the top positions in the world at a popular sci-fi game called Star Craft 2.

Unlike chess and go Star Craft 2 is a real time game and you don’t know what your opponent is doing unless you scout it. Making the game much more complicated than chess or go, AlphaStar the A.I. spent a long time looking at data (replays of professionals) to develop strategies and understanding as well as self-learning. Self-learning is where it would play against itself and see what would work what wouldn’t and how to counter different styles.

This project is the work of the DeepMind team who have been building and testing there A.I. Alpha Star, on members of the team, itself and pro-gamers. One of the main focuses of the team was trying to keep it fair, they didn’t want the computer to win because it reacted in superhuman times or because it could click insanely fast so there were limits on it so it didn’t do anything that a human couldn’t theoretically do.

To get better at the game there are 4 main points that need to be perfected:

  • Imperfect information – You do not have all the information and to get any information you need to send something to scout the map.
  • Long term planning – You need to plan ahead, how will you win in the end what can you do now to help you later?
  • Real time – Choices need to be made as fast as limited, since your opponent won’t wait for you it is always on the go.
  • Huge numbers of choices – Many buildings and units to control at once.

Post Author: Thomas Simpson