Few of my colleagues were playing foosball. And I was thinking what if they could play with computers.
“It’s not that we need a computerised foosball table, but it is a small example of a much larger problem”
Many people have asked this question to me “Is not Artificial Intelligence & Machine Learning as good as what we fed into it”
Answer is Yes and No.
Yes because Computers can learn what they are taught to learn.
No because Computers can learn from whatever they experience, they never forget and they can build neural networks of decisions on their learnings.
From my experience, there are following models of Machine Learning:
- Reinforcement Learning
Most things we would like a computer to do involves lots of decisions. And output of one decision could be input to another decision. For example, If foosball has to be automated — Ball placement, Handle movement thereby Men movement like decisions will have to be interlinked. Any time that the computer wins, we give it a reward, and any time that computer loses, we subtract a reward. To completely automate a foosball, we need to build a system which maximises rewards.
When computer plays a game again and again. It starts with choosing an action based on probability model. It has three actions-stay, move right, move left. As the computer wins and loses, it’ll strengthen its decision matrix. So during the next game, capabilities of machine increases and it becomes a better player.
- Supervised Learning
Supervised Learning works a little differently than reinforcement learning, unlike the previous model..it tags every decision as positive and negative and what type of contributions it will make in determining the output.
Machines will always start off clueless but with time and enough iteration, will get better at making decisions just by making lots and lots of decisions. This is not the the case for humans, ideally.