A Q-Table greatly simplified the challenge of helping a computer agent “learn” to solve an environment. Unfortunately, this particular approach doesn’t scale well to the kinds of applications I would like to create. To help overcome this next hurdle, we will raise the complexity a bit more as the Frozen Lake environment is approached again, this time by using a neural network.
Telling a computer to perform an action based on an input isn’t too hard. Teaching a computer to learn what action to take based on what it sees is a whole different challenge. Now imagine that the computer wont even know if the action is good or bad until some unknown point in the future – how hard would that be? Well, let’s find out as we take a look at a machine learning algorithm called Q-Learning.
Machine Learning provides us an interesting way to solve special kinds of problems. If you’re just playing around, you may see that creating a good problem to work with can be a lot of work on its own. OpenAI gym has recognized this challenge and provided a great solution. They have created a whole collection of different “environments” that are perfectly suited to machine learning. To help us get started, we will be looking at one of the easy challenges which we can solve using Q-Learning.
You may have noticed things have been quiet here recently. The reason is that I have been hard at work trying to learn new things myself. Machine Learning is to blame for my currently distracted state, but if you haven’t looked into it, perhaps I can help you catch the bug too.