Source code for gameanalysis.regret

"""A module for computing regret and social welfare of profiles"""
import itertools
import multiprocessing

import numpy as np
from scipy import optimize


[docs]def pure_strategy_deviation_gains(game, prof): """Returns the pure strategy deviations gains The result is a compact array of deviation gains. Each element corresponds to the deviation from strategy i to strategy j ordered by (i, j) for all valid deviations.""" prof = np.asarray(prof, int) supp = prof > 0 num_supp = game.role_reduce(supp) from_inds = np.arange(game.num_role_strats)[supp] reps = game.num_strategies[game.role_indices[from_inds]] num_devs = np.sum(num_supp * (game.num_strategies - 1)) to_inds = np.ones(reps.sum(), int) to_inds[0] = 0 to_inds[reps[:-1].cumsum()] -= reps[:-1] role_inds = (num_supp * game.num_strategies)[:-1].cumsum() to_inds[role_inds] += game.num_strategies[:-1] to_inds = to_inds.cumsum() to_inds = to_inds[to_inds != from_inds.repeat(reps)] from_inds = from_inds.repeat(reps - 1) pays = game.get_payoffs(prof)[from_inds] dev_profs = prof[None].repeat(num_devs, 0) dev_profs[np.arange(num_devs), from_inds] -= 1 dev_profs[np.arange(num_devs), to_inds] += 1 dev_pays = np.array([game.get_payoffs(dprof)[to] for dprof, to in zip(dev_profs, to_inds)]) return dev_pays - pays
[docs]def pure_strategy_regret(game, prof): """Returns the regret of a pure strategy profile If prof has more than one dimension, the last dimension is taken as a set of profiles and returned as a new array.""" prof = np.asarray(prof, int) return max(pure_strategy_deviation_gains(game, prof).max(), 0)
[docs]def mixture_deviation_gains(game, mix, assume_complete=False): """Returns all the gains from deviation from a mixed strategy The result is ordered by role, then strategy.""" mix = np.asarray(mix, float) strategy_evs = game.deviation_payoffs(mix, assume_complete=assume_complete) # strategy_evs is nan where there's no data, however, if it's not played in # the mix, it doesn't effect the role_evs masked = strategy_evs.copy() masked[mix == 0] = 0 role_evs = game.role_reduce(masked * mix, keepdims=True) return strategy_evs - role_evs
[docs]def mixture_regret(game, mix): """Return the regret of a mixture profile""" mix = np.asarray(mix, float) return mixture_deviation_gains(game, mix).max()
[docs]def pure_social_welfare(game, profile): """Returns the social welfare of a pure strategy profile in game""" profile = np.asarray(profile, int) return game.get_payoffs(profile).dot(profile)
[docs]def mixed_social_welfare(game, mix): """Returns the social welfare of a mixed strategy profile""" return game.get_expected_payoffs(mix).dot(game.num_players)
[docs]class SocialWelfareOptimizer(object): """A pickleable object to find Nash equilibria This method uses constrained convex optimization to to attempt to solve a proxy for the nonconvex regret minimization.""" def __init__(self, game, gtol=1e-8): self.game = game self.scale = game.max_payoffs() - game.min_payoffs() self.scale[self.scale == 0] = 1 # In case payoffs are the same self.offset = game.min_payoffs() self.gtol = gtol
[docs] def obj_func(self, mix, penalty): # pragma: no cover # We assume that the initial point is in a constant sum subspace, and # so project the gradient so that any gradient step maintains that # constant step. Thus, sum to 1 is not one of the penalty terms # Because deviation payoffs uses log space, we max with 0 just for the # payoff calculation ep, ep_jac = self.game.get_expected_payoffs( np.maximum(0, mix), assume_complete=True, jacobian=True) # Normalize so payoffs are effectively in [0, 1] ep = (ep - self.offset) / self.scale ep_jac /= self.scale[:, None] # Compute normalized negative walfare (minimization) welfare = -self.game.num_players.dot(ep) dwelfare = -self.game.num_players.dot(ep_jac) # Add penalty for negative mixtures welfare += penalty * np.sum(np.minimum(mix, 0) ** 2) / 2 dwelfare += penalty * np.minimum(mix, 0) # Project grad so steps stay in the simplex (more or less) dwelfare -= self.game.role_repeat(self.game.role_reduce(dwelfare) / self.game.num_strategies) return welfare, dwelfare
def __call__(self, mix): # pragma: no cover # Pass in lambda, and make penalty not a member result = None penalty = np.sum(self.game.num_players) for _ in range(30): # First get an unconstrained result from the optimization with np.errstate(over='raise', invalid='raise'): try: opt = optimize.minimize( lambda m: self.obj_func(m, penalty), mix, method='CG', jac=True, options={'gtol': self.gtol}) except FloatingPointError: # pragma: no cover penalty *= 2 continue mix = opt.x # Project it onto the simplex, it might not be due to the penalty result = self.game.simplex_project(mix) if np.allclose(mix, result): break # Increase constraint penalty penalty *= 2 return result
[docs]def max_mixed_social_welfare(game, grid_points=2, random_restarts=0, processes=None, **swopt_args): """Returns the maximum role symmetric mixed social welfare profile Arguments --------- grid_points : int > 1, optional The number of grid points to use for mixture seeds. two implies just pure mixtures, more will be denser, but scales exponentially with the dimension. random_restarts : int, optional The number of random initializations. processes : int, optional Number of processes to use when finding Nash equilibria. The game needs to be pickleable. """ assert game.is_complete(), \ "Max welfare finding only works on complete games""" initial_points = list(itertools.chain( [game.uniform_mixture()], game.grid_mixtures(grid_points), game.biased_mixtures(), game.role_biased_mixtures(), game.random_mixtures(random_restarts))) chunksize = len(initial_points) if processes == 1 else 4 best = (-np.inf, -1, None) opt = SocialWelfareOptimizer(game, **swopt_args) with multiprocessing.Pool(processes) as pool: for i, mix in enumerate(pool.imap_unordered( opt, initial_points, chunksize=chunksize)): welfare = mixed_social_welfare(game, mix) best = max(best, (welfare, i, mix)) return best[0], best[2]
[docs]def max_pure_social_welfare(game, by_role=False): """Returns the maximum social welfare over the known profiles. If by_role is specified, then max social welfare applies to each role independently.""" if by_role: if game.num_profiles: welfares = game.role_reduce(game.profiles * game.payoffs) prof_inds = np.nanargmax(welfares, 0) return (welfares[prof_inds, np.arange(game.num_roles)], game.profiles[prof_inds]) else: welfares = np.empty(game.num_roles) welfares.fill(np.nan) profiles = np.empty(game.num_roles, dtype=object) profiles.fill(None) return welfares, profiles else: if game.num_profiles: welfares = np.sum(game.profiles * game.payoffs, 1) prof_ind = np.nanargmax(welfares) return welfares[prof_ind], game.profiles[prof_ind] else: return np.nan, None