Two Player Wimblepong

This project is a two player version of the pong-v0 OpenAI Gym environment. The environment is used in the Reinforcement Learning course at Aalto University, Finland.

How to use

Clone the repository: https://github.com/aalto-intelligent-robotics/Wimblepong.
Install the dependencies listed in requirements.txt.
Check out the examples test_pong_ai_multi.py and test_pong_ai.py to see how to make an agent play against the included SimpleAI or two SimpleAIs against each other.

Environments

The environments as defined in __init__.py are:

WimblepongVisualMultiplayer-v0: Two agents are playing against each other based on the pixel observation
WimblepongVisualSimpleAI-v0: One agent plays against a SimpleAI based on the pixel observations
WimblepongMultiplayer-v0: Two agents are playing against each other based on the absolute positions of the ball and the paddles
WimblepongSimpleAI-v0: One agent plays against a SimpleAI based on the absolute positions of the ball and the paddles

Interface

The interface is designed to be used like the OpenAI Gym environment.

`env.step()`

Takes the input action as parameter (0: STAY, 1: UP, 2: DOWN): Takes either one action if the opponent is a SimpleAI or a tuple of two actions if two agents are playing against each other.
Returns observation: A returned observation is an array of (210, 160, 3) RGB values. If the game is played against an agent that uses the absolute values, such as SimpleAI, the returned observation is one array. If two agents that use pixel observations play against each other, the returned values is a tuple of two observations, one for each player. In each observation, the agents paddle is always green and on the left side of the arena and the opponents paddle is red and on the other side. For example, if player 2 is playing on the right side, the observation for player 2 will be flipped and red and green colors will be inverted such that playing on both sides looks the same for the agent but normal when rendering the game.
Returns the rewards for each player, either a tuple or 1 value depending on if two or one action have been passed as parameter.
Returns if an episode is done.
Returns an info dict for debug information

`env.render()`

Renders the current state of the game.

`env.reset()`

Resets the position of the player paddles and the ball, usually used when an episode is over. Launches the ball in a random direction.

`env.switch_sides()`

Allows the agents to switch sides and also switches the sides of the scoreboard. (The observations will still look the same with the agent using the green paddle no matter on which side the agent plays)

`set_names(p1, p2)`

Function to pass the agent names to the environment. The names will also be displayed on the scoreboard

SimpleAI

The SimpleAI agent is an agent that uses the absolute ball and player positions to follow the ball and reflect it in random directions.

The implementation follows Kaixhin's.

DimitrisCC / RL-Project