Playing with Reinforcement Learning

 

The article will cover topics like Reinforcement learning, OpenAI gym, RL process, World Models and MDN-RNN

Let me explain the story of RL with a example, 

lets say u are thrown into a jungle, now how would u survive?

u need to find food to survive, u need to fight carnivore animals and your need to reach a final destination where there are other humans!!!

now this example explains how RL works

this is the definition:

The jungle is the "environment", you are the "agent" and the place and situation you are in is your "game state" the food u eat is the "reward", a day in jungle is "Episode", the time is simply timestep.

and this is the process:

Agents take action in the environment which would change the state and give u a reward(+/-).

no this sounds vague, right, so how do i get a hands on experience of that, well for that u need to visit

https://www.gymlibrary.dev/index.html

Now how will our agent behave in this new environment, who will guide him, what will tell him the needed actions and how exactly is learning happening here?

well, The solution consists of 3 distinct parts

1. A variational autoencoder
2. RNN with a mixture density network
3. controller

The VAE takes the data and tells the observation 

"VAE can take a high dimensional input image and condense it into a latent random variable that approximately follows a standard multivariate normal distribution, through minimization of the reconstruction error and KL divergence.

MDN-RNN does forward thinking.

Controller is a densely connected neural network having input from both VAE and MDN-RNN.