pymoo 0.6.1.3__cp311-cp311-macosx_10_9_universal2.whl → 0.6.1.5.dev0__cp311-cp311-macosx_10_9_universal2.whl

pymoo/util/value_functions.py

@@ -0,0 +1,719 @@
+import numpy as np
+import pymoo
+import scipy
+from pymoo.optimize import minimize as moomin
+from scipy.optimize import minimize as scimin
+from pymoo.core.problem import Problem
+import matplotlib.pyplot as plt
+from scipy.optimize import NonlinearConstraint
+from scipy.optimize import Bounds
+from pymoo.algorithms.soo.nonconvex.es import ES
+from pymoo.algorithms.soo.nonconvex.ga import GA
+import math 
+from operator import mul
+from functools import reduce
+from pymoo.termination.default import DefaultSingleObjectiveTermination
+import sys
+
+
+# Input 1: A list of non-dominated points 
+# Input 2: The ranking of the given non-dominated points 
+# Input 3: constraint function for optimizing the value func
+# Input 4: the skeleton utility function that we're trying to optimize
+def create_vf(P, ranks, ineq_constr, vf="linear", delta=0.1, eps_max=1000, method="trust-constr", verbose=False): 
+
+    if vf == "linear":
+        return create_linear_vf(P, ranks, delta, eps_max, method, verbose)
+    elif vf == "poly":
+        return create_poly_vf(P, ranks, delta, eps_max, method, verbose)
+
+    else:
+        raise ValueError("Value function '%d' not supported." % vf) 
+
+    return lambda f_new:  np.sum(f_new)
+
+def create_poly_vf(P, ranks, delta=0.1, eps_max=1000, method="trust-constr", verbose=False):
+
+    if method == "trust-constr" or method == "SLSQP": 
+        return create_vf_scipy_poly(P, ranks, delta, eps_max, method=method, verbose=verbose)
+    elif method == "ES": 
+        return create_vf_pymoo_poly(P, ranks, delta, eps_max, method=method, verbose=verbose)
+    else: 
+        raise ValueError("Optimization method %s not supported" % method) 
+
+
+
+def create_linear_vf(P, ranks, delta=0.1, eps_max=1000, method="trust-constr"): 
+    
+    if method == "trust-constr" or method == "SLSQP": 
+        return create_vf_scipy_linear(P, ranks, delta, eps_max, method)
+    elif method == "ES": 
+        return create_vf_pymoo_linear(P, ranks, delta, eps_max, method)
+    else: 
+        raise ValueError("Optimization method %s not supported" % method) 
+
+
+def create_vf_scipy_poly(P, ranks, delta, eps_max, method="trust-constr", verbose=False):
+
+    # Gathering basic info
+    M = P.shape[1]
+
+    P_count = P.shape[0]
+
+    # Inequality constraints - check that each term of S in our obj is non-negative for each P
+    ineq_lb = [-np.inf] * (P_count*M)
+    ineq_ub = [0] * (P_count*M)
+
+    # Inequality constraints - check VF is monotonically increasing with user preference
+    ineq_lb += [-np.inf] * (P_count - 1)
+    ineq_ub += [0] * (P_count - 1)
+
+
+    # Equality constraints (Make sure all terms in VF add up to 1 per term of product)
+    for m in range(M): 
+        ineq_lb.append(0)
+        ineq_ub.append(0)
+
+    P_sorted = _sort_P(P, ranks)
+    ranks.sort() 
+
+    constr = NonlinearConstraint(_build_constr_poly(P_sorted, poly_vf, ranks, delta), ineq_lb, ineq_ub)
+  
+    # Bounds on x 
+    x_lb = []
+    x_ub = []
+    for m in range(M**2):
+        x_lb.append(0)
+        x_ub.append(1)
+
+    for m in range(M): 
+        x_lb.append(-1000)
+        x_ub.append(1000)
+
+    x_lb.append(-1000)
+    x_ub.append(eps_max)
+   
+    bounds = Bounds(x_lb, x_ub)
+
+    # Initial position
+    x0 = [1] * (M**2 + M + 1)
+
+    if method == 'trust-constr':
+        # The trust-constr method always finds the decision space linear
+        hess = lambda x: np.zeros((len(x0), len(x0)))
+    else: 
+        hess = None
+
+    res = scimin(_obj_func, 
+                 x0, 
+                 constraints=constr, 
+                 bounds=bounds, 
+                 method=method, 
+                 hess=hess)
+
+    # package up results 
+    vf =  lambda P_in: poly_vf(P_in, res.x[0:-1])
+    params = res.x[0:-1]
+    epsilon = res.x[-1]
+
+    fit = _validate_vf(res, verbose)
+
+    return vfResults(vf, params, epsilon, fit)
+
+
+def create_vf_scipy_linear(P, ranks, delta, eps_max, method="trust-constr", verbose=False): 
+
+    # Gathering basic info
+    M = P.shape[1]
+    
+    # Sort P
+    P_sorted = _sort_P(P, ranks)
+    ranks.sort()
+
+    # Inequality constraints
+    lb = [-np.inf] * (P.shape[0] - 1)
+    ub = [0] * (P.shape[0] - 1)
+
+    # Equality constraints
+    lb.append(0)
+    ub.append(0)
+
+    constr = NonlinearConstraint(_build_constr_linear(P_sorted, linear_vf, ranks, delta), lb, ub)
+
+    # Bounds on x
+    x_lb = []
+    x_ub = []
+
+    for m in range(M): 
+        x_lb.append(0)
+        x_ub.append(1)
+
+    x_lb.append(-1000)
+    x_ub.append(eps_max)
+   
+    bounds = Bounds(x_lb, x_ub)
+
+    # Initial position
+    x0 = [0.5] * (M+1)
+
+    if method == 'trust-constr':
+        # The trust-constr method always finds the decision space linear
+        hess = lambda x: np.zeros((len(x0), len(x0)))
+    else: 
+        hess = None
+
+    res = scimin(_obj_func, 
+                 x0, 
+                 constraints= constr,
+                 bounds=bounds, 
+                 method=method, 
+                 hess=hess)
+
+    # package up results
+    vf =  lambda P_in: linear_vf(P_in, res.x[0:-1])
+    params = res.x[0:-1]
+    epsilon = res.x[-1]
+    fit = _validate_vf(res, verbose)
+
+    return vfResults(vf, params, epsilon, fit)
+
+def create_vf_pymoo_linear(P, ranks, delta, eps_max, method="ES", verbose=False): 
+
+    vf_prob = OptimizeLinearVF(P, ranks, delta, eps_max, linear_vf)
+
+    if method == "ES":
+        algorithm = ES()
+    elif method == "GA": 
+        algorithm = GA()
+    else: 
+        raise ValueError("Optimization method %s not supported" % method) 
+
+    res = moomin(vf_prob,
+        algorithm,
+        ('n_gen', 200),
+        verbose=verbose,
+        seed=1)
+
+    vf = lambda P_in: linear_vf(P_in, res.X[0:-1])
+
+    if res.X is not None:
+        params = res.X[0:-1]
+        epsilon = res.X[-1]
+    else: 
+        params = None
+        epsilon = -1000
+
+    fit = _validate_vf(res, verbose)
+
+    return vfResults(vf, params, epsilon, fit)
+
+
+def create_vf_pymoo_poly(P, ranks, delta, eps_max, method="trust-constr", verbose=False):
+
+    vf_prob = OptimizePolyVF(P, ranks, delta, eps_max, poly_vf)
+
+    if method == "ES":
+        algorithm = ES()
+    elif method == "GA": 
+        algorithm = GA()
+    else: 
+        raise ValueError("Optimization method %s not supported" % method) 
+
+    res = moomin(vf_prob,
+        algorithm,
+        ('n_gen', 100),
+        verbose=verbose,
+        seed=1)
+
+    vf = lambda P_in: poly_vf(P_in, res.X[0:-1])
+
+    if res.X is not None:
+        params = res.X[0:-1]
+        epsilon = res.X[-1]
+    else: 
+        params = None
+        epsilon = -1000
+
+    fit = _validate_vf(res, verbose)
+
+    return vfResults(vf, params, epsilon, fit)
+
+
+def linear_vf(P, x): 
+
+    return np.matmul(P, x.T)
+
+def poly_vf(P, x): 
+
+    # find out M
+    M = P.shape[-1]
+
+    result = []
+
+    if len(x.shape) == 1:
+        x_len = 1    
+    else:
+        x_len = x.shape[0]
+
+    # Calculate value for each row of x 
+    for xi in range(x_len):
+
+        running_product = 1
+
+        # Get current x 
+        if x_len == 1:
+            curr_x = x
+        else: 
+            curr_x = x[xi, :]
+
+        S = _calc_S(P, curr_x)
+
+        # Multiply all of the S terms together along the final axis 
+        product = np.prod(S, axis = len(np.shape(S))-1)
+
+        result.append(product)
+
+    return np.squeeze(np.array(result))
+
+
+def plot_vf(P, vf, show=True): 
+
+    plt.scatter(P[:,0], P[:,1], marker=".", color="red", s=200 )
+  
+
+    for i in range(np.size(P,0)):
+        plt.annotate("P%d" % (i+1), (P[i,0], P[i,1]))
+
+    min_x = min(P[:,0])
+    min_y = min(P[:,1])
+    max_x = max(P[:,0])
+    max_y = max(P[:,1])
+
+    x,y = np.meshgrid(np.linspace(min_x, max_x, 1000), np.linspace(min_y, max_y, 1000))
+
+    z = vf(np.stack((x,y), axis=2))
+
+    values_at_P = []
+    for p in range(np.size(P,0)):
+        values_at_P.append(vf(P[p,:]))
+
+    values_at_P.sort()
+
+    plt.contour(x,y,z, levels=values_at_P)
+
+    plt.colorbar()
+
+    if show: 
+        plt.show()
+
+    return plt
+
+
+
+## ---------------- Polynomial VF creation functions ------------------
+
+def _ineq_constr_poly(x, P, vf, ranks, delta):
+    if len(x.shape) == 1:
+        return _ineq_constr_1D_poly(x, P, vf, ranks, delta)
+    else: 
+        return _ineq_constr_2D_poly(x, P, vf, ranks, delta)
+
+
+def _build_ineq_constr_poly(P, vf, ranks, delta):
+
+    ineq_func = lambda x : _ineq_constr_poly(x, P, vf, ranks, delta)
+
+    return ineq_func
+
+def _eq_constr_poly(x):
+
+    M = math.floor(math.sqrt(6))
+
+    if len(x.shape) == 1:
+        result = [] 
+        for m in range(M): 
+            result.append(-(sum(x[m*M:m*M+M]) - 1))
+
+    else: 
+
+        pop_size = np.size(x,0)
+
+        result = []
+
+        for xi in range(pop_size): 
+
+            result_for_xi = [] 
+
+            for m in range(M): 
+
+                result_for_xi.append(-(sum(x[xi,m*M:m*M+M]) - 1))
+
+            result.append(result_for_xi)
+
+        result = np.array(result)
+
+    return result
+
+
+def _build_constr_poly(P, vf, ranks, delta): 
+
+    ineq_constr_func = _build_ineq_constr_poly(P, vf, ranks, delta)
+
+    return lambda x : np.append(ineq_constr_func(x), _eq_constr_poly(x))
+
+
+def _ineq_constr_2D_poly(x, P, vf, ranks, delta):
+
+    pop_size = np.size(x,0)
+
+    P_count = P.shape[0]
+    M = P.shape[1]
+
+    S_constr_len = M * P_count 
+    increasing_len = P_count - 1
+
+    G = np.ones((pop_size, S_constr_len + increasing_len))*-99
+   
+    for xi in range(pop_size): 
+    
+        G[xi, :] = _ineq_constr_1D_poly(x[xi, :], P, vf, ranks, delta)
+
+    return G
+
+
+def _ineq_constr_1D_poly(x, P, vf, ranks, delta):
+
+    ep = x[-1]
+
+    P_count = P.shape[0]
+    M = P.shape[1]
+
+    S_constr_len = M * P_count 
+    increasing_len = P_count - 1
+
+    G = np.ones((1, S_constr_len + increasing_len))*-99
+
+    # Checking to make sure each S term in the polynomial objective function is non-negative
+    S = _calc_S(P, x[0:-1]) * -1
+
+    G[:, 0:S_constr_len] = S.reshape(1, S_constr_len)
+
+
+    # Pair-wise compare each ranked member of P, seeing if our proposed utility 
+    #  function increases monotonically as rank increases
+    for p in range(P_count - 1):
+
+        current_P_val = vf(P[[p],:], x[0:-1])
+        next_P_val = vf(P[[p+1],:], x[0:-1])
+
+        current_rank = ranks[p]
+        next_rank = ranks[p+1]
+
+        if current_rank == next_rank: 
+            # Handle ties
+            G[:,[p + S_constr_len]] = np.abs(current_P_val - next_P_val) - delta*ep
+        else: 
+            G[:,[p + S_constr_len]] = -(current_P_val - next_P_val) + ep
+
+
+    return G
+
+## ---------------- Linear VF creation functions ------------------
+
+def _build_ineq_constr_linear(P, vf, ranks, delta):
+
+    ineq_func = lambda x : _ineq_constr_linear(x, P, vf, ranks, delta)
+
+    return ineq_func
+
+def _build_constr_linear(P, vf, ranks, delta):
+
+    ineq_constr_func = _build_ineq_constr_linear(P, vf, ranks, delta);
+
+    return lambda x : np.append(ineq_constr_func(x), _eq_constr_linear(x))
+
+
+def _ineq_constr_linear(x, P, vf, ranks, delta):
+    if len(x.shape) == 1:
+        return _ineq_constr_1D_linear(x, P, vf, ranks, delta)
+    else: 
+        return _ineq_constr_2D_linear(x, P, vf, ranks, delta)
+
+
+def _ineq_constr_2D_linear(x, P, vf, ranks, delta):
+
+    ep = np.column_stack([x[:,-1]]) 
+    pop_size = np.size(x,0)
+
+    G = np.ones((pop_size, np.size(P,0)-1))*-99
+
+    # Pair-wise compare each ranked member of P, seeing if our proposed utility 
+    #  function increases monotonically as rank increases
+    for p in range(np.size(P,0) - 1):
+
+
+        current_P_val = vf(P[[p],:], x[:, 0:-1])
+        next_P = vf(P[[p+1],:], x[:, 0:-1])
+
+        current_rank = ranks[p]
+        next_rank = ranks[p+1]
+
+        
+        if current_rank == next_rank: 
+            # Handle ties 
+            G[:,[p]] = np.abs(current_P_val.T - next_P.T) - delta*ep
+        else: 
+            # As vf returns, each column is an value of P for a given x in the population
+            # We transpose to make each ROW the value of P
+            G[:,[p]] = -(current_P_val.T - next_P.T) + ep
+
+
+    return G
+
+
+def _ineq_constr_1D_linear(x, P, vf, ranks, delta):
+
+    ep = x[-1]
+
+    G = np.ones((1, np.size(P,0)-1))*-99
+
+    # Pair-wise compare each ranked member of P, seeing if our proposed utility 
+    #  function increases monotonically as rank increases
+    for p in range(np.size(P,0) - 1):
+
+        current_P_val = vf(P[[p],:], x[0:-1])
+        next_P = vf(P[[p+1],:], x[0:-1])
+
+        current_rank = ranks[p]
+        next_rank = ranks[p+1]
+
+        if current_rank == next_rank: 
+            # Handle ties 
+            G[:,[p]] = np.abs(current_P_val - next_P) - delta*ep
+        else: 
+            G[:,[p]] = -(current_P_val - next_P) + ep
+        
+
+
+    return G
+
+def _obj_func(x): 
+
+    # check if x is a 1D array or a 2D array
+    if len(x.shape) == 1:
+        ep = x[-1]
+    else: 
+        ep = np.column_stack([x[:,-1]])
+
+    return -ep
+
+def _sort_P(P, ranks): 
+    P_with_rank = np.hstack((P, np.array([ranks]).T))
+
+    P_sorted = P[P_with_rank[:, -1].argsort()]
+
+    return P_sorted
+
+# Can have several P instances, but assumes one x instance 
+def _calc_S(P, x): 
+
+    M = P.shape[-1]
+
+    # reshape x into a matrix called k
+    k = np.array(x[0:M*M].reshape(M,M), copy=True)
+    k = k.T  
+   
+    # roll each column down to work with equation 5
+    for col in range(1, k.shape[1]):
+        k[:,[col]] = np.roll(k[:,[col]], col)
+
+
+    l = x[M*M:(M*M)+M].reshape(M,1)
+    l = l.T
+
+    # Calc S for an arbitrary dimensional space of P 
+    S = np.matmul(P, k) + (l  * 2)
+
+    return np.squeeze(S)
+
+
+# Constraint that states that the x values must add up to 1
+def _eq_constr_linear(x): 
+
+    if len(x.shape) == 1:
+        eq_cons = sum(x[0:-1]) - 1
+    else: 
+        eq_cons = np.sum(x[:,0:-1],1, keepdims=True) - 1
+
+    return eq_cons
+
+
+# Makes a comparator for a given value function and the P that is ranked second in the 
+def make_vf_comparator(vf, P_rank_2):
+
+    return lambda P : vf_comparator(vf, P_rank_2, P)
+
+def vf_comparator(vf, P_rank_2, P):
+
+    reference_value = vf(P_rank_2)
+
+    if reference_value > vf(P):
+        return -1
+    elif reference_value < vf(P): 
+        return 1
+    else: 
+        return 0 
+
+
+class OptimizeLinearVF(Problem): 
+
+    def __init__(self, P, ranks, delta, eps_max, vf):
+       
+        # One var for each dimension of the object space, plus epsilon 
+        n_var_vf = np.size(P, 1) + 1
+
+        # it has one inequality constraints for every pair of solutions in P
+        n_ieq_c_vf = np.size(P,0) - 1
+       
+        xl_vf = [0.0] * n_var_vf 
+        xu_vf = [1.0] * n_var_vf
+        
+        # upper/lower bound on the epsilon variable is -1000/1000
+        xl_vf[-1] = -1000
+        xu_vf[-1] = eps_max
+
+        # TODO start everything at 0.5
+
+        self.P = _sort_P(P, ranks)
+
+        self.ranks = ranks
+        self.ranks.sort()
+
+        self.vf = vf
+
+        self.ranks = ranks
+        self.delta = delta
+
+        super().__init__(n_var_vf, n_obj=1, n_ieq_constr=n_ieq_c_vf, n_eq_constr=1, xl=xl_vf, xu=xu_vf)
+
+    def _evaluate(self, x, out, *args, **kwargs):
+
+        ## Objective function: 
+        obj = _obj_func(x)
+
+        # The objective function above returns a negated version of epsilon 
+        ep = -obj
+
+        # maximize epsilon, or the minimum distance between each contour 
+        out["F"] = obj
+
+        ## Inequality
+
+        ineq_func = _build_ineq_constr_linear(self.P, self.vf, self.ranks, self.delta)
+
+        out["G"] = ineq_func(x)
+            
+        ## Equality constraint that keeps sum of x under 1
+        out["H"] = _eq_constr_linear(x)
+
+def _validate_vf(res, verbose):
+
+    message = "" 
+
+    if isinstance(res, pymoo.core.result.Result):
+        success = np.all(res.G <= 0)
+        epsilon = res.X[-1]
+        if not success: 
+            message = "Constraints not met\n"
+        if epsilon < 0:
+            message = message + "Epsilon negative\n"
+
+
+    elif isinstance(res, scipy.optimize.optimize.OptimizeResult):
+        success = res.success and res.constr_violation <= 0
+        epsilon = res.x[-1]
+
+        if not (res.constr_violation <= 0): 
+            message = "Constraints not met."
+        else:
+            message = res.message
+    else: 
+        ValueError("Internal error: bad result objective given for validation")
+
+    if epsilon < 0 or not success: 
+    
+        if verbose:
+            sys.stderr.write("WARNING: Unable to fit value function\n")    
+            sys.stderr.write(message + "\n")    
+
+        return False
+    else:
+        return True
+
+
+class OptimizePolyVF(Problem): 
+
+    def __init__(self, P, ranks, delta, eps_max, vf):
+      
+        M = P.shape[1]
+
+        P_count = P.shape[0]
+
+
+        # One var for each dimension of the object space, plus epsilon 
+        n_var_vf = (M**2 + M + 1)
+
+        # it has one inequality constraints for every pair of solutions in P
+        n_ieq_c_vf =  (P_count*M) + (P_count - 1)
+       
+        xl_vf = [0.0] * n_var_vf 
+        xu_vf = [1.0] * n_var_vf
+        
+        # upper/lower bound on the epsilon variable is -1000/1000
+        xl_vf[-1] = -1000
+        xu_vf[-1] = eps_max 
+
+        # TODO start everything at 0.5
+
+        self.P = _sort_P(P, ranks)
+
+        self.ranks = ranks
+        self.ranks.sort()
+
+        self.vf = vf
+
+        self.delta = delta
+
+        super().__init__(n_var_vf, n_obj=1, n_ieq_constr=n_ieq_c_vf, n_eq_constr=M, xl=xl_vf, xu=xu_vf)
+
+    def _evaluate(self, x, out, *args, **kwargs):
+
+        ## Objective function: 
+        obj = _obj_func(x)
+
+        # The objective function above returns a negated version of epsilon 
+        ep = -obj
+
+        # maximize epsilon, or the minimum distance between each contour 
+        out["F"] = obj
+
+        ## Inequality
+        ineq_func = _build_ineq_constr_poly(self.P, self.vf, self.ranks, self.delta)
+
+        out["G"] = ineq_func(x)
+            
+        ## Equality constraint that keeps sum of x under 1
+        out["H"] = _eq_constr_poly(x)
+
+
+class vfResults(): 
+
+    def __init__(self, vf, params, epsilon, fit): 
+
+        self.vf = vf
+        self.params = params
+        self.epsilon = epsilon  
+        self.fit = fit 
+
+
+