Robocup

RoboCup 2D Soccer Simulation League, written in Python, for my A.I. class project.

In case you're wondering, this won the class tournament, but its positioning is shitty tho.

Installation

Clone this repo:

git clone <git-url>

Assuming you have rcsoccersim on your machine, run it first. Then run the AI agent optionally with a teamname prepended.

rcsoccersim
[TEAM=<NAME>] python main.py

Development

aigent/aima_python/ is just for reference - that's the sample code from Norvig's Artificial Intelligence, A Modern Approach.

aigent/soccerpy/ contains the source for the agent and client. This is sourced from jsontbradshaw/soccerpy under the MIT license. I've modified it to add several convenient methods, but minimal modification within this folder is needed.

aigent/agent_*.py is the real deal - it lets you configure different agent personalities and stategies. For example you can have striker, defender, goalie agents, which are agent_1, agent_2 and agent_3 respectively.

Once you have the agents, you can import a distribution of them, assign them player numbers and configure their starting positions in main.py. This is also the file you run to play.

Report

Read the paper at report/paper.pdf

Notes

Below are my notes containing strategies I used, and ideas I had no time to implement.

kickoff

act if it's your team's ball. is_before_kick_off() && is_kick_off_us()

whistle

Throw in, free kick, penalty, goalkick BEFORE_KICK_OFF = "before_kick_off" PLAY_ON = "play_on" TIME_OVER = "time_over" KICK_OFF_L = "kick_off_l" KICK_OFF_R = "kick_off_r" KICK_IN_L = "kick_in_l" KICK_IN_R = "kick_in_r" FREE_KICK_L = "free_kick_l" FREE_KICK_R = "free_kick_r" CORNER_KICK_L = "corner_kick_l" CORNER_KICK_R = "corner_kick_r" GOAL_KICK_L = "goal_kick_l" GOAL_KICK_R = "goal_kick_r" DROP_BALL = "drop_ball" OFFSIDE_L = "offside_l" OFFSIDE_R = "offside_r"

offense

When your team has the ball.

advance: if not close enuf
shoot: if close enuf and clear

defense

When ball is in your half of the field, and enemy's possession

just swarm them for now

player maneuver

block, evades

Neural Net

on top of predefined conditional actions and heuristics

Fields, per robot want: input = conditions, output = actions

env input:

distance to goal
distance to ball
distance to enemy
distance to teammate

output actions:

shoot to goal (scoring is the whole point of the game)
pass to teammate
run to position (ball, teammate, enemy, side)

unsupervised pretraining on autoencoder:

output = input, i.e. input env conds

supervised training:

set {env, actions} collected in the last round of game, only if result is positive

First define ur env vars n simple heuristics

or skip the pretraining by using a random net first, plug in net n repeat

run a game, each think step collect data of outcome n prev actions, inject to db. game ends, train autoencoder w/ all actions, then train real plug in the net, repeat above for evolution

where to incorp action success: reward n cost? in selection of data, by evolution n natural selection