Compare various agents
"""
import os
from typing import List
import io
from urllib.request import urlopen
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
from rlcard3.games.mocsar.agentdb import str_to_agent_dict, get_by_id
from rlcard3.utils.config_read import Config
conf = Config('environ.properties')
# PATH Const
LOG_SAVE_PRFX = conf.get_str(section='cfg.compare', key="stat_dir_path")
PNG_SAVE_PRFX = conf.get_str(section='cfg.visual', key="png_dir_path")
log_dirname = conf.get_str(section='cfg.visual', key="dir_name")
log_filename = conf.get_str(section='cfg.visual', key="file_name")
def read_data_local() -> pd.DataFrame:
csv_file_name = os.path.join(LOG_SAVE_PRFX, log_dirname, log_filename)
dfr = pd.read_csv(csv_file_name, sep=";", usecols=["cardnr", "agentid", "agentstr", "payoff"])
return dfr
def read_data_github(csv_url:str) -> pd.DataFrame:
r1 = urlopen(csv_url)
df1 = pd.read_csv(io.BytesIO(r1.read()),
compression='gzip',
import rlcard3
from rlcard3.games.mocsar.stat import MocsarStat
from rlcard3.utils.config_read import Config
from rlcard3.utils.utils import tournament
conf = Config('environ.properties')
NR_GAMES = conf.get_int(section='cfg.compare', key='nr_games')
# Make environment and enable human mode
env = rlcard3.make('mocsar-cfg', config={'multi_agent_mode': True})
# Create statistics
stat = MocsarStat(game=env.game,
agents=env.model.rule_agents,
nr_of_games=NR_GAMES,
batch_name=conf.get_str(section='cfg.compare',
key='batch_name'),
log_dir=conf.get_str(section='cfg.compare',
key='stat_dir_path'))
# Register agents
agent_str = conf.get_str(section='cfg.compare', key="agent_str")
nr_cards = conf.get_int(section='global', key='nr_cards')
agents = {agent_str: 1, "mocsar_random": 3}
print(f"mocsar_pl_dqn_pytorch_load_model_cfg, Agents:{agents}")
# # Here we directly load NFSP models from /models module
# rl_agents = models.load(agent_str,
# num_players=env.game.get_player_num(),
# action_num=env.action_num,
import rlcard3
from rlcard3.games.mocsar.agentdb import str_to_agent_list
from rlcard3.games.mocsar.stat import MocsarStat
from rlcard3.utils.config_read import Config
conf = Config('environ.properties')
NR_GAMES = conf.get_int(section='cfg.compare', key='nr_games')
# Make environment and enable human mode
env = rlcard3.make('mocsar-cfg', config={'multi_agent_mode': True})
# Create statistics
stat = MocsarStat(game=env.game,
agents=env.model.rule_agents,
nr_of_games=NR_GAMES,
batch_name=conf.get_str(section='cfg.compare',
key='batch_name'),
log_dir=conf.get_str(section='cfg.compare',
key='stat_dir_path'))
# Register agents
agents_list = str_to_agent_list(agent_str_list=conf.get_str(section='cfg.compare', key="agent_list"))
print(f"mocsar_pl_cfg_config, Agents:{agents_list}")
# Try different agent combinations
# List of nr of cards
card_nr_list = [i for i in range(15, 56, 4)] + [i for i in range(62, 111, 4)]
for agents in agents_list:
env.model.create_agents(agents)
print(f"Agents:{[(ag.__str__() + ', ') for ag in env.model.rule_agents]}")
# Iterate over nr of cards
for nr_cards in card_nr_list:
env.game.set_game_params(num_players=4, num_cards=nr_cards)
''' Another example of loading a pre-trained NFSP model on Leduc Hold'em
Here, we directly load the model from model zoo
'''
import rlcard3
from rlcard3.agents.random_agent import RandomAgent
from rlcard3.utils.utils import set_global_seed, tournament
from rlcard3 import models
from rlcard3.utils.config_read import Config
# Make environment
env = rlcard3.make('mocsar')
# Get parameters
conf = Config('environ.properties')
evaluate_num = conf.get_int(section='cfg.compare', key='nr_games')
agent_str = conf.get_str(section='cfg.compare', key="agent_str")
nr_cards = conf.get_int(section='global', key='nr_cards')
# Set a global seed
#set_global_seed(0)
# Here we directly load NFSP models from /models module
dqn_agents = models.load(agent_str,
num_players=env.game.get_player_num(),
action_num=env.action_num,
state_shape=env.state_shape).agents
# Evaluate the performance. Play with random agents.
random_agent = RandomAgent(env.action_num)
env.game.set_game_params(num_players=4, num_cards=nr_cards)
env.set_agents([dqn_agents[0], random_agent, random_agent, random_agent])