rugby like simulation played using policy gradient method

ATTACKING AND DEFENDING TEAMS: 3 players

• DEF: plays randomly
• ATT: single actor(ball bearer) or whole team can learn

VICTORY CONDITIONS:

• time expires -> DEF
• ball passes try line -> ATT

FIELD:

• by default 13(long)* 19(wide)
• try line: 1/3 of the field (0,0 at upper left corner)

ACTIONS:

• advance (motion toward opponent endline)
• step back (toward its endline; also un-stuck the player)
• dodge (lateral left/right)
• pass the ball (finite maximum distance; also available back pass only rule)
• tackle (stops a opponent or steals the ball, if tackled has it)

POLICY

• policy gradient based (REINFORCE)
• softmax in actions preferences to describe policy mapping

NETWORK [ keras.layers.Flatten(FIELD), keras.layers.Dense(32, activation="softmax"), keras.layers.Dense( activation='softmax',ACTIONS_NUMBER), keras.layers.Reshape(ACTIONS{either single or multiple actors}) ]

REWARDS

• game lost: -200,
• time step passed: -5,
• lost stolen by defense team: -5,
• game won: 200,
• advancing with the ball: +20,
• average game distance too big for ball pass: -40,
• ball stolen from defense team: 0,
• accomplishing ball pass: 0,
• keeping the position:-40

OPTIMIZER

• Adam
• using gradient ascent

• field display: opencv (tested 4.6.0)
• loss dispay: matplotlib(tested 3.5.2), tkinter NOTE: if unavailable, will only throw a warning
• training: tensorflow (tested 2.9.1)

{python3} scripts/decision.py [-h] help

[-s] single actor: only ball bearer learns; default False

[-w WIDTH] field width; default 9

[-l LENGHT] field lenght; default 13

[-d DURATION] game duration in steps; default 20

[-g GAMES] number of games to play; default 500

[-v] will print actions performed by players in terminal; default False

[-t] will print field state in terminal; default False

[-r] will display field image; default False

[-f FRAME_DURATION] the lower, the higher the render frame rate; default 10ms

[--RNG] plays a completely random match (no learning or loss printing)

Ctrl+C stops the simulation and prints partial results

scripts/rugby_players.py: Player class definition, distane and average team distance functions

scripts/rugby_game.py: game definition, rules, actions, displaying

scripts/decision.py: launches the game, learning agent definition, argparser