DataPredict

Getting Started With Reinforcement Learning

Requirements

What Is Reinforcement Learning?

Reinforcement learning is a way for our models to learn on its own without the labels.

We can expect our models to perform poorly at the start of the training but they will gradually improve over time.

The Basics

Environment Feature Vector

An environment feature vector is a vector containing all the information related to model’s environment. It can contain as many information such as:

An example of environment feature vector will look like this:

local environmentFeatureVector = {

  {1, -32, 234, 12, -97} -- 1 is added at first column for bias, but it is optional.

}

Reward Value

This is the value where we reward or punish the models. The properties of reward value is shown below:

It is recommended to set the reward that is within the range of:

-1 <= (total reward * learning rate) <= 1

Action Labels

Action label is a label produced by the model. This label can be a part of decision-making classes or classification classes. For example:

Setting Up Our Reinforcement Learning Model

There are two ways we can setup our models for our environment:

Below we will show you the difference between the two above. But first, let’s define a number of variables and setup the first part of the model.


local ClassesList = {1, 2}

local NeuralNetwork = DataPredict.Models.NeuralNetwork.new() -- Create the NeuralNetwork first.

NeuralNetwork:addLayer(4, true, "None")

NeuralNetwork:addLayer(2, false, "LeakyReLU")

NeuralNetwork:setClassesList(ClassesList)

local DeepQLearning = DataPredict.Models.DeepQLearning.new() -- Then create the DeepQLearning.

DeepQLearning:setModel(NeuralNetwork) -- Then put the NeuralNetwork inside DeepQLearning.

Classic Setup

All the reinforcement learning models have two important functions:

Below, I will show a code sample using these functions.


while true do

  local previousEnvironmentFeatureVector = 0 -- We must keep track our previous feature vector.

  local action = 1

  for step = 1, 1000, 1 do

    local environmentVector = fetchEnvironmentFeatureVector(previousEnvironmentVector, action)

    action = DeepQLearning:predict(environmentVector)[1][1]

    local reward = getReward(environmentVector)

    DeepQLearning:categoricalUpdate(previousEnvironmentVector, reward, action, environmentVector) -- update() is called whenever a step is made.

    previousEnvironmentFeatureVector = environmentVector

    local hasGameEnded = checkIfGameHasEnded(environmentVector)

    if hasGameEnded then break end

  end

  QLearningNeuralNetwork:episodeUpdate() -- episodeUpdate() is used whenever an episode ends. An episode is the total number of steps that determines when the model should stop training.

end

As you can see, there are a lot of things that we must track of, but it gives you total freedom on what you want to do with the reinforcement learning models.

Quick Setup

To reduce the amount of things we need to track, we can use CategoricalPolicy in “QuickSetups” section.


local DeepQLearningQuickSetup = DataPredict.QuickSetups.CategoricalPolicy.new()

DeepQLearningQuickSetup:setModel(DeepQLearning)

DeepQLearningQuickSetup:setClassesList(classesList)

local environmentFeatureVector = 0

local action = 1

while true do

  environmentFeatureVector = fetchEnvironmentFeatureVector(environmentFeatureVector, action)

  local reward = getReward(environmentFeatureVector, action)

  action = DeepQLearningQuickSetup:reinforce(environmentFeatureVector, reward)

end

As you can see, the CategoricalPolicy compresses a number of codes into reinforce() function.

Wrapping It All Up

In this tutorial, you have learnt the starting point of the reinforcement learning neural networks.

These only cover the basics. You can find more information here: