[DO NOT MERGE] Testing get_data_dir

quantecon/game_theory/__init__.py

@@ -4,6 +4,8 @@
 
 """
 from .normal_form_game import Player, NormalFormGame
+from .polymatrix_game import PolymatrixGame
+
 from .normal_form_game import pure2mixed, best_response_2p
 from .random import (
     random_game, covariance_game, random_pure_actions, random_mixed_actions
@@ -13,6 +15,8 @@
 from .lemke_howson import lemke_howson
 from .mclennan_tourky import mclennan_tourky
 from .vertex_enumeration import vertex_enumeration, vertex_enumeration_gen
+from .howson_lcp import polym_lcp_solver
+
 from .game_generators import (
     blotto_game, ranking_game, sgc_game, tournament_game, unit_vector_game
 )
@@ -21,3 +25,5 @@
 from .localint import LocalInteraction
 from .brd import BRD, KMR, SamplingBRD
 from .logitdyn import LogitDynamics
+
+from .game_converters import qe_nfg_from_gam_file

quantecon/game_theory/game_converters.py

@@ -0,0 +1,51 @@
+"""
+Functions for converting between ways of storing games.
+"""
+
+from .normal_form_game import NormalFormGame
+from itertools import product
+
+
+def qe_nfg_from_gam_file(filename):
+    """
+    Makes a QuantEcon Normal Form Game from a gam file.
+    Gam files are described by GameTracer.
+    http://dags.stanford.edu/Games/gametracer.html
+
+    Args:
+        filename: path to gam file.
+
+    Returns:
+        NormalFormGame:
+            The QuantEcon Normal Form Game
+            described by the gam file.
+    """
+    with open(filename, 'r') as file:
+        lines = file.readlines()
+        combined = [
+            token
+            for line in lines
+            for token in line.split()
+        ]
+
+        i = iter(combined)
+        players = int(next(i))
+        actions = [int(next(i)) for _ in range(players)]
+
+        nfg = NormalFormGame(actions)
+
+        entries = [
+                {
+                    tuple(reversed(action_combination)): float(next(i))
+                    for action_combination in product(
+                        *[range(a) for a in actions])
+                }
+                for _ in range(players)
+            ]
+
+        for action_combination in product(*[range(a) for a in actions]):
+            nfg[tuple(reversed(action_combination))] = [
+                entries[p][action_combination] for p in range(players)
+                ]
+
+    return nfg

quantecon/game_theory/howson_lcp.py

@@ -0,0 +1,124 @@
+import numpy as np
+from .polymatrix_game import PolymatrixGame
+from ..optimize.pivoting import _pivoting, _lex_min_ratio_test
+from ..optimize.lcp_lemke import _get_solution
+
+
+def polym_lcp_solver(polym: PolymatrixGame):
+    """
+    Finds the Nash Equilbrium of a polymatrix game
+    using Howson's algorithm described in
+    https://www.jstor.org/stable/2634798
+    which utilises linear complimentarity.
+
+    Args:
+        polym (PolymatrixGame): Polymatrix game to solve.
+
+    Returns:
+        Probability distribution across actions for each player
+        at Nash Equilibrium.
+    """
+    LOW_AVOIDER = 2.0
+    # makes all of the costs greater than 0
+    positive_cost_maker = polym.range_of_payoffs()[1] + LOW_AVOIDER
+    # Construct the LCP like Howson:
+    M = np.vstack([
+        np.hstack([
+            np.zeros(
+                (polym.nums_actions[player], polym.nums_actions[player])
+                ) if p2 == player
+            else (positive_cost_maker - polym.polymatrix[(player, p2)])
+            for p2 in range(polym.N)
+        ] + [
+            -np.outer(np.ones(
+                polym.nums_actions[player]), np.eye(polym.N)[player])
+        ])
+        for player in range(polym.N)
+    ] + [
+        np.hstack([
+            np.hstack([
+                np.outer(np.eye(polym.N)[player], np.ones(
+                    polym.nums_actions[player]))
+                for player in range(polym.N)
+            ]
+            ),
+            np.zeros((polym.N, polym.N))
+        ])
+    ]
+    )
+    total_actions = sum(polym.nums_actions)
+    q = np.hstack([np.zeros(total_actions), -np.ones(polym.N)])
+
+    n = np.shape(M)[0]
+    tableau = np.hstack([
+        np.eye(n),
+        -M,
+        np.reshape(q, (-1, 1))
+    ])
+
+    basis = np.array(range(n))
+    z = np.empty(n)
+
+    starting_player_actions = {
+        player: 0
+        for player in range(polym.N)
+    }
+
+    for player in range(polym.N):
+        row = sum(polym.nums_actions) + player
+        col = n + sum(polym.nums_actions[:player]) + \
+            starting_player_actions[player]
+        _pivoting(tableau, col, row)
+        basis[row] = col
+
+    # Array to store row indices in lex_min_ratio_test
+    argmins = np.empty(n + polym.N, dtype=np.int_)
+    p = 0
+    retro = False
+    while p < polym.N:
+        finishing_v = sum(polym.nums_actions) + n + p
+        finishing_x = n + \
+            sum(polym.nums_actions[:p]) + starting_player_actions[p]
+        finishing_y = finishing_x - n
+
+        pivcol = (
+            finishing_v if not retro
+            else finishing_x if finishing_y in basis
+            else finishing_y
+        )
+
+        retro = False
+
+        while True:
+
+            _, pivrow = _lex_min_ratio_test(
+                tableau, pivcol, 0, argmins,
+            )
+
+            _pivoting(tableau, pivcol, pivrow)
+            basis[pivrow], leaving_var = pivcol, basis[pivrow]
+
+            if leaving_var == finishing_x or leaving_var == finishing_y:
+                p += 1
+                break
+            elif leaving_var == finishing_v:
+                print("entering the backtracking case")
+                p -= 1
+                retro = True
+                break
+            elif leaving_var < n:
+                pivcol = leaving_var + n
+            else:
+                pivcol = leaving_var - n
+
+    combined_solution = _get_solution(tableau, basis, z)
+
+    eq_strategies = [
+        combined_solution[
+            sum(polym.nums_actions[:player]):
+            sum(polym.nums_actions[:player+1])
+            ]
+        for player in range(polym.N)
+    ]
+
+    return eq_strategies

quantecon/game_theory/polymatrix_game.py

@@ -0,0 +1,138 @@
+import numpy as np
+from itertools import product
+from .normal_form_game import NormalFormGame, Player
+
+
+def hh_payoff_player(
+        nf: NormalFormGame,
+        my_player_number,
+        my_action_number,
+        is_polymatrix=False
+):
+    """
+    hh stands for head-to-head.
+    Approximates the payoffs to a player when they play
+    an action with values at the actions of other players
+    that try to sum to the payoff.
+    Precise when the game can be represented with a polymatrix.
+
+    Args:
+        my_player_number: The number of the player making the action.
+        my_action_number: The number of the action.
+        is_polymatrix:
+            Is the game represented by this normal form
+            actually a polymatrix game.
+
+    Returns:
+        Dictionary giving an approximate component
+        of the payoff at each action for each other player.
+    """
+    action_combinations = product(*(
+        [
+            range(nf.nums_actions[p]) if p != my_player_number
+            else [my_action_number]
+            for p in range(nf.N)]
+    ))
+    my_player: Player = nf.players[my_player_number]
+    my_payoffs = my_player.payoff_array
+    hh_actions_and_payoffs = np.vstack([
+        np.hstack(
+            [
+                np.eye(nf.nums_actions[p])[action_combination[p]]
+                for p in range(nf.N) if p != my_player_number
+            ] + [
+                my_payoffs[
+                    action_combination[my_player_number:]
+                    + action_combination[:my_player_number]
+                ]
+            ]
+        )
+        for action_combination in action_combinations
+    ])
+    hh_actions = hh_actions_and_payoffs[:, :-1]
+    combined_payoffs = hh_actions_and_payoffs[:, -1]
+
+    # different ways to solve the simultaneous equations
+    if is_polymatrix:
+        # this does not go much faster and
+        # could also be used for games with no actual polymatrix
+        hh_payoffs_array, residuals, _, _ = np.linalg.lstsq(
+            hh_actions, combined_payoffs, rcond=None
+        )
+        assert not residuals, "The game is not actually a polymatrix game"
+    else:
+        hh_payoffs_array = np.dot(np.linalg.pinv(hh_actions), combined_payoffs)
+
+    payoff_labels = [
+        (p, a)
+        for p in range(nf.N) if p != my_player_number
+        for a in range(nf.nums_actions[p])
+    ]
+
+    payoffs = {label: payoff for label, payoff in zip(
+        payoff_labels, hh_payoffs_array)}
+
+    return payoffs
+
+
+class PolymatrixGame:
+    """
+    In a polymatrix game, the payoff to a player is the sum
+    of their payoffs from bimatrix games against each player.
+    i.e. If two opponents deviate, the change in payoff
+    is the sum of the changes in payoff of each deviation.
+    """
+    def __str__(self) -> str:
+        str_builder = ""
+        for k, v in self.polymatrix.items():
+            str_builder += str(k) + ":\n"
+            str_builder += str(v) + "\n\n"
+        return str_builder
+
+    def __init__(self, number_of_players, nums_actions, polymatrix) -> None:
+        self.N = number_of_players
+        self.nums_actions = nums_actions
+        self.polymatrix = polymatrix
+
+    @classmethod
+    def from_nf(cls, nf: NormalFormGame, is_polymatrix=False):
+        """
+        Creates a Polymatrix approximation to a
+        Normal Form Game. Precise if possible.
+
+        Args:
+            nf (NormalFormGame): Normal Form Game to approximate.
+            is_polymatrix :
+                Is the game represented by the normal form
+                actually a polymatrix game.
+
+        Returns:
+            New Polymatrix Game.
+        """
+        polymatrix_builder = {
+            (p1, p2): np.full(
+                (nf.nums_actions[p1], nf.nums_actions[p2]), -np.inf)
+            for p1 in range(nf.N)
+            for p2 in range(nf.N)
+            if p1 != p2
+        }
+        for p1 in range(nf.N):
+            for a1 in range(nf.nums_actions[p1]):
+                payoffs = hh_payoff_player(
+                    nf, p1, a1, is_polymatrix=is_polymatrix)
+                for ((p2, a2), payoff) in payoffs.items():
+                    polymatrix_builder[(p1, p2)][a1][a2] = payoff
+
+        return cls(nf.N, nf.nums_actions, polymatrix_builder)
+
+    def range_of_payoffs(self):
+        """
+        The lowest and highest components of payoff from
+        head to head games.
+
+        Returns:
+            Tuple of minimum and maximum.
+        """
+        min_p = min([np.min(M) for M in self.polymatrix.values()])
+        max_p = max([np.max(M) for M in self.polymatrix.values()])
+        return (min_p, max_p)

quantecon/game_theory/tests/test_howson_lcp.py

@@ -0,0 +1,34 @@
+"""
+Tests for howson_lcp.py
+"""
+
+from numpy.testing import assert_
+from quantecon.game_theory.game_converters import qe_nfg_from_gam_file
+from quantecon.game_theory.howson_lcp import polym_lcp_solver
+from quantecon.game_theory.polymatrix_game import PolymatrixGame
+
+import os
+
+
+# Mimicing quantecon.tests.util.get_data_dir
+data_dir_name = "gam_files"
+this_dir = os.path.dirname(__file__)
+data_dir = os.path.join(this_dir, data_dir_name)
+
+
+def test_polym_lcp_solver_where_solution_is_pure_NE():
+    filename = "big_polym.gam"
+    nfg = qe_nfg_from_gam_file(os.path.join(data_dir, filename))
+    polym = PolymatrixGame.from_nf(nfg)
+    ne = polym_lcp_solver(polym)
+    worked = nfg.is_nash(ne)
+    assert_(worked)
+
+
+def test_polym_lcp_solver_where_lcp_solver_must_backtrack():
+    filename = "triggers_back_case.gam"
+    nfg = qe_nfg_from_gam_file(os.path.join(data_dir, filename))
+    polym = PolymatrixGame.from_nf(nfg)
+    ne = polym_lcp_solver(polym)
+    worked = nfg.is_nash(ne)
+    assert_(worked)