NAPRL(Not Another Platform for Reinforcement Learning)

NAPRL is a free library for developping Reinforcement Learning algorithm. It is written in python and use pygame for rendering the game. It contains of environment you can use for testing implementing your artificial intelligence algorithm. Ii is mainly aimed to use Reinforcement Learning algorithm but the architecture is very flexible and you can use other artificial intelligence technique like Goal Orientef Behaviour,Fuzzy logic,etc...

There are so many Reinforcement platforms... Why using NAPRL ?

NAPRL differs from the other platform by his simplicity and his ready to go testing. Also many of the RL platform use tensorflow as for implementing the algorithm. Our platform will be more focused on pytorch for deep learning algorithm. You do not need to install many depedencies to run it. You just need to install python and pygame. If you want to train your agent using Deep Q Network, install pytorch. In addition,you can create easily your environment and customize the state and action of the environment.

Installation

Installing numpy

pip install numpy

Installing pygame

pip install pygame

Installing pytorch

pip install torch

Architecture

Environment

The environment defines the rules of game.For now,we have included one environment which is the famous grid world. Each environment hiils a number of agent which interact with it.I fyou want too create your own environment, just inherit from this class .

Check environment.py for more details.

Agent

The agent is the entity who interacts with the environment. The agent is controlled by a brain who is reponsible for his decision. This class is abstract and if you have to inherit from it when you create new environment.

Brain

The brain is the learning algorithm you attach to your agent decision making. Currently,the engines support two brain which are the Q Learning and DeepQNetwork. This is the main component interface you have define if you have to implemted your learning algorithm. See brain.py to check how it was implemented. The Q Learning and the Deep Q Network algorithms inherit from the brain interface

Using Q Learning algorithm with the gridworld environment

You can check main.py for more details of the implementation

Below are the step.

import your environment

We import the gridworld environment. Check Environment/gridworld/gridworld_environment.py for more details.

env = GD_Environment(screen)

Choose your learning algorithm (Q Learning or DeepQNetwork )

Here we choose Q Learning if you want to choose Q Learning:

env.setMainAgent(Brain_Type.Q_Learning)

If you want to choose DeepQ Network

env.setMainAgent(Brain_Type.Deep_Q_Network)

Make your model in training mode

env.training =True

Define the number of episodes

See Machine_Learning/hyper_parameters.py for some defined constants like the number of episodes

num_episodes = NUM_EPISODE

Train your model

max_frames = 200

for i in range(num_episodes):

    done = False

    env.reset(i)

    env.main_agent.entity.x,env.main_agent.entity.y = 32 * random.randint(1,10),32 *random.randint(1,10)

    env.render()

    frame = 0

    time.sleep(0.01)
    print("Episode ",i)
    while not done:

        frame+=1
        done = env.step()

        env.render()

        if frame >=max_frames:
            done =True
        time.sleep(0.01)