Source code for axelrod.strategies.hmm

from typing import Any, Dict

from axelrod.action import Action
from axelrod.evolvable_player import (
    EvolvablePlayer,
    InsufficientParametersError,
    copy_lists,
    crossover_lists,
)
from axelrod.player import Player

C, D = Action.C, Action.D


[docs]def is_stochastic_matrix(m, ep=1e-8) -> bool: """Checks that the matrix m (a list of lists) is a stochastic matrix.""" for i in range(len(m)): for j in range(len(m[i])): if (m[i][j] < 0) or (m[i][j] > 1): return False s = sum(m[i]) if abs(1.0 - s) > ep: return False return True
def normalize_vector(vec): s = sum(vec) vec = [v / s for v in vec] return vec
[docs]def mutate_row(row, mutation_probability, rng): """, crossover_lists_of_lists Given a row of probabilities, randomly change each entry with probability `mutation_probability` (a value between 0 and 1). If changing, then change by a value randomly (uniformly) chosen from [-0.25, 0.25] bounded by 0 and 100%. """ randoms = rng.random(len(row)) for i in range(len(row)): if randoms[i] < mutation_probability: ep = rng.uniform(-1, 1) / 4 row[i] += ep if row[i] < 0: row[i] = 0 if row[i] > 1: row[i] = 1 return row
[docs]class SimpleHMM(object): """Implementation of a basic Hidden Markov Model. We assume that the transition matrix is conditioned on the opponent's last action, so there are two transition matrices. Emission distributions are stored as Bernoulli probabilities for each state. This is essentially a stochastic FSM. https://en.wikipedia.org/wiki/Hidden_Markov_model """ def __init__( self, transitions_C, transitions_D, emission_probabilities, initial_state ) -> None: """ Params ------ transitions_C and transitions_D are square stochastic matrices: lists of lists with all values in [0, 1] and rows that sum to 1. emission_probabilities is a vector of values in [0, 1] initial_state is an element of range(0, len(emission_probabilities)) """ self.transitions_C = transitions_C self.transitions_D = transitions_D self.emission_probabilities = emission_probabilities self.state = initial_state self._cache_C = dict() # type: Dict[int, int] self._cache_D = dict() # type: Dict[int, int] self._cache_deterministic_transitions() # Random generator will be set by parent strategy self._random = None # type: Any def _cache_deterministic_transitions(self): """Cache deterministic transitions to avoid unnecessary random draws.""" # If 1 is in the transition vector, it's deterministic. Just pick it out. # By caching we avoid repeated searches. for state in range(len(self.transitions_C)): if 1 in self.transitions_C[state]: self._cache_C[state] = self.transitions_C[state].index(1) for state in range(len(self.transitions_D)): if 1 in self.transitions_D[state]: self._cache_D[state] = self.transitions_D[state].index(1)
[docs] def is_well_formed(self) -> bool: """ Determines if the HMM parameters are well-formed: - Both matrices are stochastic - Emissions probabilities are in [0, 1] - The initial state is valid. """ if not is_stochastic_matrix(self.transitions_C): return False if not is_stochastic_matrix(self.transitions_D): return False for p in self.emission_probabilities: if (p < 0) or (p > 1): return False if self.state not in range(0, len(self.emission_probabilities)): return False return True
def __eq__(self, other: Player) -> bool: """Equality of two HMMs""" check = True for attr in [ "transitions_C", "transitions_D", "emission_probabilities", "state", ]: check = check and getattr(self, attr) == getattr(other, attr) return check
[docs] def move(self, opponent_action: Action) -> Action: """Changes state and computes the response action. Parameters opponent_action: Axelrod.Action The opponent's last action. """ # Choose next state. if opponent_action == C: try: next_state = self._cache_C[self.state] except KeyError: num_states = len(self.emission_probabilities) next_state = self._random.choice(num_states, 1, p=self.transitions_C[self.state])[0] else: try: next_state = self._cache_D[self.state] except KeyError: num_states = len(self.emission_probabilities) next_state = self._random.choice(num_states, 1, p=self.transitions_D[self.state])[0] self.state = next_state # Choose action to emit. p = self.emission_probabilities[self.state] if p == 0: return D if p == 1: return C action = self._random.random_choice(p) return action
[docs]class HMMPlayer(Player): """ Abstract base class for Hidden Markov Model players. Names - HMM Player: Original name by Marc Harper """ name = "HMM Player" classifier = { "memory_depth": 1, "stochastic": True, "long_run_time": False, "inspects_source": False, "manipulates_source": False, "manipulates_state": False, } def __init__( self, transitions_C=None, transitions_D=None, emission_probabilities=None, initial_state=0, initial_action=C ) -> None: Player.__init__(self) if not transitions_C: transitions_C = [[1]] transitions_D = [[1]] emission_probabilities = [0.5] # Not stochastic initial_state = 0 self.initial_state = initial_state self.initial_action = initial_action self.hmm = SimpleHMM( copy_lists(transitions_C), copy_lists(transitions_D), list(emission_probabilities), initial_state ) assert self.hmm.is_well_formed() self.state = self.hmm.state self.classifier["stochastic"] = self.is_stochastic()
[docs] def is_stochastic(self) -> bool: """Determines if the player is stochastic.""" # If the transitions matrices and emission_probabilities are all 0 or 1 # Then the player is stochastic values = set(self.hmm.emission_probabilities) for m in [self.hmm.transitions_C, self.hmm.transitions_D]: for row in m: values.update(row) if not values.issubset({0, 1}): return True return False
[docs] def strategy(self, opponent: Player) -> Action: if len(self.history) == 0: return self.initial_action else: action = self.hmm.move(opponent.history[-1]) # Record the state for testing purposes, this isn't necessary # for the strategy to function self.state = self.hmm.state return action
[docs] def set_seed(self, seed=None): super().set_seed(seed=seed) # Share RNG with HMM # The evolvable version of the class needs to manually share the rng with the HMM # after initialization. try: self.hmm._random = self._random except AttributeError: pass
[docs]class EvolvableHMMPlayer(HMMPlayer, EvolvablePlayer): """Evolvable version of HMMPlayer.""" name = "EvolvableHMMPlayer" def __init__( self, transitions_C=None, transitions_D=None, emission_probabilities=None, initial_state=0, initial_action=C, num_states=None, mutation_probability=None, seed: int = None ) -> None: EvolvablePlayer.__init__(self, seed=seed) transitions_C, transitions_D, emission_probabilities, initial_state, initial_action, num_states, mutation_probability = self._normalize_parameters( transitions_C, transitions_D, emission_probabilities, initial_state, initial_action, num_states, mutation_probability) self.mutation_probability = mutation_probability HMMPlayer.__init__(self, transitions_C=transitions_C, transitions_D=transitions_D, emission_probabilities=emission_probabilities, initial_state=initial_state, initial_action=initial_action) self.hmm._random = self._random self.overwrite_init_kwargs( transitions_C=transitions_C, transitions_D=transitions_D, emission_probabilities=emission_probabilities, initial_state=initial_state, initial_action=initial_action, num_states=num_states, mutation_probability=mutation_probability ) def _normalize_parameters(self, transitions_C=None, transitions_D=None, emission_probabilities=None, initial_state=None, initial_action=None, num_states=None, mutation_probability=None): if not ((transitions_C and transitions_D and emission_probabilities) and (initial_state is not None) and (initial_action is not None)): if not num_states: raise InsufficientParametersError("Insufficient Parameters to instantiate EvolvableHMMPlayer") transitions_C, transitions_D, emission_probabilities, initial_state, initial_action = self.random_params( num_states) # Normalize types of various matrices for m in [transitions_C, transitions_D]: for i in range(len(m)): m[i] = list(map(float, m[i])) emission_probabilities = list(map(float, emission_probabilities)) num_states = len(emission_probabilities) if mutation_probability is None: mutation_probability = 10 / (num_states ** 2) else: mutation_probability = mutation_probability return transitions_C, transitions_D, emission_probabilities, initial_state, initial_action, num_states, mutation_probability def random_params(self, num_states): transitions_C = [] transitions_D = [] emission_probabilities = [] for _ in range(num_states): transitions_C.append(self._random.random_vector(num_states)) transitions_D.append(self._random.random_vector(num_states)) emission_probabilities.append(self._random.random()) initial_state = self._random.randint(0, num_states) initial_action = C return transitions_C, transitions_D, emission_probabilities, initial_state, initial_action @property def num_states(self): return len(self.hmm.emission_probabilities) def mutate_rows(self, rows, mutation_probability): for i, row in enumerate(rows): row = mutate_row(row, mutation_probability, self._random) rows[i] = normalize_vector(row) return rows
[docs] def mutate(self): transitions_C = self.mutate_rows( self.hmm.transitions_C, self.mutation_probability) transitions_D = self.mutate_rows( self.hmm.transitions_D, self.mutation_probability) emission_probabilities = mutate_row( self.hmm.emission_probabilities, self.mutation_probability, self._random) initial_action = self.initial_action if self._random.random() < self.mutation_probability / 10: initial_action = self.initial_action.flip() initial_state = self.initial_state if self._random.random() < self.mutation_probability / (10 * self.num_states): initial_state = self._random.randint(0, self.num_states) return self.create_new( transitions_C=transitions_C, transitions_D=transitions_D, emission_probabilities=emission_probabilities, initial_state=initial_state, initial_action=initial_action, )
[docs] def crossover(self, other): if other.__class__ != self.__class__: raise TypeError("Crossover must be between the same player classes.") transitions_C = crossover_lists(self.hmm.transitions_C, other.hmm.transitions_C, self._random) transitions_D = crossover_lists(self.hmm.transitions_D, other.hmm.transitions_D, self._random) emission_probabilities = crossover_lists( self.hmm.emission_probabilities, other.hmm.emission_probabilities, self._random) return self.create_new( transitions_C=transitions_C, transitions_D=transitions_D, emission_probabilities=emission_probabilities )
[docs] def receive_vector(self, vector): """ Read a serialized vector into the set of HMM parameters (less initial state). Then assign those HMM parameters to this class instance. Assert that the vector has the right number of elements for an HMMParams class with self.num_states. Assume the first num_states^2 entries are the transitions_C matrix. The next num_states^2 entries are the transitions_D matrix. Then the next num_states entries are the emission_probabilities vector. Finally the last entry is the initial_action. """ assert(len(vector) == 2 * self.num_states ** 2 + self.num_states + 1) def deserialize(vector): matrix = [] for i in range(self.num_states): row = vector[self.num_states * i: self.num_states * (i + 1)] row = normalize_vector(row) matrix.append(row) return matrix break_tc = self.num_states ** 2 break_td = 2 * self.num_states ** 2 break_ep = 2 * self.num_states ** 2 + self.num_states initial_state = 0 self.hmm = SimpleHMM( deserialize(vector[0:break_tc]), deserialize(vector[break_tc:break_td]), normalize_vector(vector[break_td:break_ep]), initial_state ) self.initial_action = C if round(vector[-1]) == 0 else D self.initial_state = initial_state
[docs] def create_vector_bounds(self): """Creates the bounds for the decision variables.""" vec_len = 2 * self.num_states ** 2 + self.num_states + 1 lb = [0.0] * vec_len ub = [1.0] * vec_len return lb, ub
[docs]class EvolvedHMM5(HMMPlayer): """ An HMM-based player with five hidden states trained with an evolutionary algorithm. Names: - Evolved HMM 5: Original name by Marc Harper """ name = "Evolved HMM 5" classifier = { "memory_depth": 5, "stochastic": True, "long_run_time": False, "inspects_source": False, "manipulates_source": False, "manipulates_state": False, } def __init__(self) -> None: initial_state = 3 initial_action = C t_C = [ [1, 0, 0, 0, 0], [0, 1, 0, 0, 0], [0, 1, 0, 0, 0], [0.631, 0, 0, 0.369, 0], [0.143, 0.018, 0.118, 0, 0.721], ] t_D = [ [0, 1, 0, 0, 0], [0, 0.487, 0.513, 0, 0], [0, 0, 0, 0.590, 0.410], [1, 0, 0, 0, 0], [0, 0.287, 0.456, 0.146, 0.111], ] emissions = [1, 0, 0, 1, 0.111] super().__init__(t_C, t_D, emissions, initial_state, initial_action)