def test_agent(): actions = [-1, 0, 1, 2, 3, 4] agent = Agent(player_index=0, actions=actions) game_state = State(n_players=3) private_state = agent.get_private_state(game_state) assert len(private_state) == agent.len_private_state assert private_state[0] == GameStage.PRE_FLOP assert private_state[1] == game_state.hole_cards[ agent.player_index][0].rank assert private_state[2] == game_state.hole_cards[ agent.player_index][0].suit assert private_state[3] == game_state.hole_cards[ agent.player_index][1].rank assert private_state[4] == game_state.hole_cards[ agent.player_index][1].suit model_input = agent.get_model_input(private_state, actions) assert model_input.shape == (len(actions), agent.n_inputs) assert np.allclose(model_input[:, -1], actions) assert np.allclose(model_input[:, 0], private_state[0]) model_input_single_action = agent.get_model_input(private_state, actions=[-1]) assert model_input_single_action.shape == (1, agent.n_inputs)
def test_minimum_legal_bet(): initial_dealer = 0 state = State(n_players=3) # Note: the first person to act is the player after the big blind, # and their minimum legal bet is therefore the big blind amount assert state.minimum_legal_bet() == state.big_blind state.update(-1) min_legal_bet = state.minimum_legal_bet() assert min_legal_bet == (state.big_blind - state.small_blind) state.update(min_legal_bet) # Note: the big blind is allowed to bet zero dollars (to "knock") assert state.minimum_legal_bet() == 0
def describe_learned_q_function(self, n_iter=20): for _ in range(n_iter): game_state = State(n_players=3) private_cards = game_state.hole_cards[self.player_index] private_state = self.get_private_state(game_state) model_input = self.get_model_input(private_state, self.actions) q = self.model.predict(model_input)[:, 0] print( f"Value at stage {game_state.game_stage.name} with private cards {private_cards}: {q}" )
def get_bot(name, hole, possible_actions, state): """Return the bot based on its name""" parsed_state = State(name, hole, possible_actions, state) bot = SmartBot(parsed_state) if 'simple' == name: bot = SimpleBot(parsed_state) elif 'random' == name: bot = RandomBot(parsed_state) elif 'threshold' == name: bot = ThresholdBot(parsed_state) elif 'agressive-loose' == name: bot = AgressiveLooseBot(parsed_state) elif 'agressive-tight' == name: bot = AgressiveTightBot(parsed_state) elif 'passive-loose' == name: bot = PassiveLooseBot(parsed_state) elif 'passive-tight' == name: bot = PassiveTightBot(parsed_state) elif 'smart' == name: bot = SmartBot(parsed_state) elif 'randomized-smart' == name: bot = RandomizedSmartBot(parsed_state) LOG.info("\n\nname of the bot: %s", bot.name) return bot
def run_one_episode(episode, players, initial_wealth=100): # Note: the probability of random (exploratory) actions decreases over time proba_random_action = 0.02 + 0.98 * np.exp(-episode / 500) state = State(n_players=len(players), initial_wealth=initial_wealth, verbose=False) learning_player = state.current_player private_state = players[learning_player].get_private_state(state) action = players[learning_player].get_action(state, proba_random_action) cumulative_reward = 0 # Note: at the beginning of every episode, we push updates to the q function # from the learning player to all other players for player_index in range(state.n_players): if player_index != learning_player: players[player_index].model.set_weights( players[learning_player].model.get_weights()) while not state.terminal: wealth_before_action = state.wealth[learning_player] state.update(action) # Note: we need to make the other players act so that we can get back to the learning player # Even though the other (non-learning) players are acting, it # is possible that the learning player is forced # to act (bet the small blind or big blind) in this block while state.current_player != learning_player: action = players[state.current_player].get_action( state, proba_random_action) state.update(action) # TODO Is this needed? if state.terminal: break # Note: we calculate the player's reward _after_ the other players act # TODO This can create odd rewards when the other players fold and # the learning player is either small or big blind in the next round wealth_after_action = state.wealth[learning_player] reward = wealth_after_action - wealth_before_action cumulative_reward += reward # print(f" *** Reward for player {learning_player} is ${reward} ***") # TODO Put this in pytest assert state.wealth[ learning_player] == initial_wealth + cumulative_reward next_private_state = players[learning_player].get_private_state(state) next_action = players[learning_player].get_action( state, proba_random_action) continuation_value = players[learning_player].predicted_q( next_private_state, next_action) if state.terminal: print( f"Reached a terminal state: player wealths are {state.wealth}") updated_guess_for_q = reward else: updated_guess_for_q = reward + continuation_value players[learning_player].update_q(private_state, action, updated_guess_for_q) action = next_action private_state = next_private_state
def test_state_with_low_wealth(): initial_wealth = 5 big_blind = 2 small_blind = 1 # Note: cards are dealt (popped) off of the _end_ of the list deck = [ Card(Rank.KING, Suit.HEARTS), Card(Rank.KING, Suit.DIAMONDS), Card(Rank.KING, Suit.CLUBS), Card(Rank.THREE, Suit.HEARTS), Card(Rank.NINE, Suit.CLUBS), Card(Rank.TWO, Suit.SPADES), Card(Rank.SEVEN, Suit.SPADES), Card(Rank.TWO, Suit.CLUBS), Card(Rank.SEVEN, Suit.CLUBS), Card(Rank.TWO, Suit.DIAMONDS), Card(Rank.SEVEN, Suit.DIAMONDS), Card(Rank.ACE, Suit.HEARTS), Card(Rank.ACE, Suit.DIAMONDS), ] state = State( n_players=4, initial_wealth=initial_wealth, big_blind=big_blind, small_blind=small_blind, deck=deck, ) state.update(big_blind) state.update(big_blind) # Note: the small blind completes state.update(big_blind - small_blind) assert state.game_stage == GameStage.FLOP for _ in range(state.n_players): state.update(small_blind) assert state.game_stage == GameStage.TURN for _ in range(state.n_players): state.update(small_blind) assert state.game_stage == GameStage.RIVER for _ in range(state.n_players): # Note: the players all knock (bet zero) state.update(0) # Note: player index 0 had a pair of aces as their hole cards, and they win the round. The next # player to be big blind has only $1 left (and the big blind is $2), so they are forced to go all in # as part of the blind assert state.wealth[0] == initial_wealth + (state.n_players - 1) * ( big_blind + 2 * small_blind ) assert ( state.wealth[2] == state.wealth[3] == initial_wealth - (big_blind + 2 * small_blind) == 1 ) assert state.game_stage == GameStage.PRE_FLOP state.update(small_blind) state.update(small_blind) assert state.game_stage == GameStage.FLOP assert state.dealer == 1 assert state.current_player == 2 # Note: the low wealth players are all in, so bets > 0 are not allowed assert state.maximum_legal_bet() == 0 for _ in range(state.n_players): state.update(0) assert state.game_stage == GameStage.TURN state.update(0) for _ in range(state.n_players - 1): state.update(-1) # Note: player 2 wins because everyone else has folded assert state.wealth[2] == 1 + small_blind * (state.n_players - 1) assert state.wealth[3] == 0
def test_state(): initial_wealth = 200 initial_dealer = 0 state = State( n_players=3, initial_wealth=initial_wealth, initial_dealer=initial_dealer ) n_cards_in_deck_after_initial_deal = 52 - state.n_players * 2 assert len(state.shuffled_deck) == n_cards_in_deck_after_initial_deal assert state.game_stage == GameStage.PRE_FLOP assert len(state.public_cards) == 0 assert all(player_has_folded is False for player_has_folded in state.has_folded) # Note: any action < 0 means the current player is folding action_fold = -1 # Note: player index 1 is the small blind, player index 2 is the big blind, # so player index 0 (the dealer) is the first to act pre flop assert state.current_player == state.dealer state.update(action_fold) assert state.has_folded[state.dealer] assert state.current_player == 1 ten_dollars = 10 complete_blind_plus_ten = (state.big_blind - state.small_blind) + ten_dollars state.update(complete_blind_plus_ten) assert state.bets_by_stage[GameStage.PRE_FLOP][1] == [ state.small_blind, complete_blind_plus_ten, ] assert state.game_stage == GameStage.PRE_FLOP # Note: the big blind player calls. At this point, # the dealer has folded, and the other two have both bet the big blind plus $10. # We move to the next stage (the flop) and the dealer deals three public cards state.update(ten_dollars) assert state.bets_by_stage[GameStage.PRE_FLOP][2] == [state.big_blind, ten_dollars] assert state.game_stage == GameStage.FLOP assert len(state.public_cards) == 3 assert len(state.shuffled_deck) == n_cards_in_deck_after_initial_deal - 3 # Note: the player after the dealer is the first to act after the flop assert state.current_player == 1 # Note: the second player bets $10, then the dealer bets $20, # the first player has already folded and does nothing, and then the second player calls state.update(ten_dollars) state.update(2 * ten_dollars) state.update(ten_dollars) # Note: the dealer folded before making any bets assert state.bets_by_stage[GameStage.FLOP][state.dealer] == [] assert state.bets_by_stage[GameStage.FLOP][1] == [ten_dollars, ten_dollars] assert state.bets_by_stage[GameStage.FLOP][2] == [2 * ten_dollars] assert state.game_stage == GameStage.TURN assert len(state.public_cards) == 4 assert state.current_player == 1 state.update(ten_dollars) assert state.current_player == 2 state.update(action_fold) total_bets_by_player_2 = ten_dollars * 3 + state.big_blind assert state.wealth[2] == initial_wealth - total_bets_by_player_2 assert state.wealth[1] == initial_wealth + total_bets_by_player_2 assert state.wealth[0] == initial_wealth assert sum(state.wealth) == initial_wealth * state.n_players assert len(state.public_cards) == 0 assert len(state.shuffled_deck) == n_cards_in_deck_after_initial_deal # Note: the dealer shifts to the next player assert state.dealer == 1 assert all(player_has_folded is False for player_has_folded in state.has_folded) assert sum(state.wealth) == initial_wealth * state.n_players action_knock = 0 action_small_bet = 10 action_big_bet = 20 assert state.game_stage == GameStage.PRE_FLOP # Note: all players bet (or complete) the big blind, and we move to the next stage state.update(state.big_blind) state.update(state.small_blind) state.update(action_knock) assert state.game_stage == GameStage.FLOP assert len(state.public_cards) == 3 for _ in range(state.n_players): # Note: all players bet $10 after the flop # and we move to the next stage state.update(action_small_bet) assert state.game_stage == GameStage.TURN assert len(state.public_cards) == 4 for _ in range(state.n_players): # Note: all players bet $10 after the turn # and we move to the next stage state.update(action_small_bet) assert state.game_stage == GameStage.RIVER assert len(state.public_cards) == 5 # Note: even though the stage has not ended, all public cards have # been dealt and we can calculate hand strengths hand_strengths, hand_descriptions = state.calculate_best_hand_strengths() winning_players = argmax(hand_strengths) # Note: there may be multiple winning players (a tie) # We assert on the first winning player's wealth wealth_before_winning = state.wealth[winning_players[0]] # Note: no players have folded, so all hand strengths must be positive assert max(hand_strengths) > 0 state.update(action_small_bet) state.update(action_big_bet) state.update(action_big_bet) # Note: this player initially made a small bet, and the next person raised them with a large bet # The first person to act now completes their bet (calls the raise) state.update(action_big_bet - action_small_bet) # Note: the bets were small during the flop and turn, and big during the river amount_won = ( (state.n_players - len(winning_players)) * (2 * action_small_bet + action_big_bet + state.big_blind) / len(winning_players) ) assert state.wealth[winning_players[0]] == wealth_before_winning + amount_won
def test_ties(): initial_wealth = 200.0 initial_dealer = 0 # Note: there will be a full house (aces and kings) on the board # Cards are dealt (popped) off of the _end_ of the list deck = [ Card(Rank.ACE, Suit.HEARTS), Card(Rank.ACE, Suit.DIAMONDS), Card(Rank.KING, Suit.HEARTS), Card(Rank.KING, Suit.DIAMONDS), Card(Rank.KING, Suit.CLUBS), Card(Rank.TWO, Suit.HEARTS), Card(Rank.SEVEN, Suit.HEARTS), Card(Rank.TWO, Suit.SPADES), Card(Rank.SEVEN, Suit.SPADES), Card(Rank.TWO, Suit.CLUBS), Card(Rank.SEVEN, Suit.CLUBS), ] # Note: we pass in a deck so that the order in which cards are dealt is known and deterministic state = State( n_players=3, initial_wealth=initial_wealth, initial_dealer=initial_dealer, deck=deck, ) action_bet = 1 action_fold = -1 action_knock = 0 state.update(state.big_blind) state.update(state.small_blind) state.update(action_knock) assert state.game_stage == GameStage.FLOP for _ in range(state.n_players): state.update(action_bet) assert state.game_stage == GameStage.TURN for _ in range(state.n_players): state.update(action_bet) assert state.game_stage == GameStage.RIVER amount_bet_by_player_who_folds = ( sum(state.bets_by_stage[0][state.current_player]) + sum(state.bets_by_stage[1][state.current_player]) + sum(state.bets_by_stage[2][state.current_player]) ) state.update(action_fold) assert state.has_folded[1] # Note: the other two players stay in the game. They tie and split the pot state.update(action_bet) state.update(action_bet) amount_won_by_each_winner = amount_bet_by_player_who_folds / (state.n_players - 1) assert state.wealth[0] == initial_wealth + amount_won_by_each_winner assert state.wealth[2] == initial_wealth + amount_won_by_each_winner assert sum(state.wealth) == initial_wealth * state.n_players