def create_problem(self, task_list):
# prepare problem data (TODO spostare calcolo di task_positions in classe funzione di costo)
self.task_list = task_list
self.n_tasks = self.guidance.n_agents
task_positions = np.array(
[np.array(t.coordinates) for t in task_list.tasks])
task_indices = [t.id for t in task_list.tasks]
starting_position = self.guidance.current_pose.position[:-1]
# create problem matrices
c = self._generate_cost(task_positions, starting_position)
A, b = self._generate_constraints(task_indices)
# create problem object
x = Variable(len(self.task_list.tasks))
obj = c @ x
constr = A.transpose() @ x == b
problem = LinearProblem(objective_function=obj, constraints=constr)
self.agent.problem = problem
# set communicator label
self.guidance.communicator.current_label = int(task_list.label)
# create algorithm object
self.algorithm = DistributedSimplex(
self.agent, stop_iterations=self.stop_iterations)
def initialize_scenario(self,
robot_id: int,
prediction_horizon: int,
system_matrices: dict,
cost_matrices: dict,
coupling_constraints: dict,
local_constraints: dict = None):
# save information
self.prediction_horizon = prediction_horizon
self.system_matrices = system_matrices
self.coupling_constraints = coupling_constraints
self.robot_id = robot_id
# shorhands
A = system_matrices['A']
B = system_matrices['B']
T = prediction_horizon
E = local_constraints['x_matrix']
f = local_constraints['x_vector']
G = local_constraints['u_matrix']
h = local_constraints['u_vector']
Q = cost_matrices['state']
R = cost_matrices['input']
n, m = self._sizes = (A.shape[0], B.shape[1])
# create optimization variables
z = Variable((T + 1) * n + T * m) # complete optimization variable
self._x = np.vstack((np.eye((T + 1) * n), np.zeros(
(T * m,
(T + 1) * n)))) @ z # state portion of optimization variable
self._u = np.vstack((np.zeros(((T + 1) * n, T * m)), np.eye(
T * m))) @ z # input portion of optimization variable
# initialize objective function and constraints
self._obj_func = 0
self._basic_constraints = []
for t in range(T):
# extract optimization variables corresponding to x(t), u(t), x(t+1)
x_t = np.eye(n, (T + 1) * n, t * n).T @ self._x
u_t = np.eye(m, T * m, t * m).T @ self._u
x_tp = np.eye(n, (T + 1) * n, (t + 1) * n).T @ self._x
# build local constraints
self._basic_constraints.append(x_tp == A.T @ x_t +
B.T @ u_t) # dynamics
self._basic_constraints.append(E.T @ x_t <= f) # state
self._basic_constraints.append(G.T @ u_t <= h) # input
# add terms to objective function
self._obj_func += QuadraticForm(x_t, Q) + QuadraticForm(u_t, R)
# Generate problem data
#####################
# points
points = np.zeros((dim, nsamp[0] + nsamp[1]))
points[:, 0:nsamp[0]] = np.random.multivariate_normal(mu[0], sigma[0],
nsamp[0]).transpose()
points[:, nsamp[0]:] = np.random.multivariate_normal(mu[1], sigma[1],
nsamp[1]).transpose()
# labels
labels = np.ones((sum(nsamp), 1))
labels[nsamp[0]:] = -labels[nsamp[0]:]
# cost function
z = Variable(dim + 1 + NN)
A = np.zeros((dim + 1 + NN, dim))
A[0:dim:, :] = np.eye(dim) # w = A @ z
B = np.zeros((dim + 1 + NN, 1))
B[dim + 1:dim + NN + 1] = np.ones((NN, 1)) # xi_1 + ... + xi_N = B @ z
D = np.zeros((dim + 1 + NN, 1))
D[dim] = 1 # b = D @ z
E = np.zeros((dim + 1 + NN, 1))
E[dim + 1 + agent_id] = 1 # xi_i = E @ z
obj_func = (1 / 2) * (A @ z) @ (A @ z) + C * (B @ z)
# constraints
F = np.zeros((dim + 1 + NN, NN))
F[dim + 1:dim + NN + 1:, :] = np.eye(NN)
from disropt.functions import QuadraticForm, Variable
from disropt.utils.graph_constructor import MPIgraph
from disropt.problems import Problem
# generate communication graph (everyone uses the same seed)
comm_graph = MPIgraph('random_binomial', 'metropolis')
agent_id, in_nbrs, out_nbrs, in_weights, _ = comm_graph.get_local_info()
# size of optimization variable
n = 5
# generate quadratic cost function
np.random.seed()
Q = np.random.randn(n, n)
Q = Q.transpose() @ Q
x = Variable(n)
func = QuadraticForm(x - np.random.randn(n, 1), Q)
# create Problem and Agent
agent = Agent(in_nbrs, out_nbrs, in_weights=in_weights)
agent.set_problem(Problem(func))
# run the algorithm
x0 = np.random.rand(n, 1)
algorithm = GradientTracking(agent, x0)
algorithm.run(iterations=1000, stepsize=0.01)
print("Agent {} - solution estimate: {}".format(
agent_id,
algorithm.get_result().flatten()))
# Generate problem data
#####################
# points
points = np.zeros((dim, nsamp[0] + nsamp[1]))
points[:, 0:nsamp[0]] = np.random.multivariate_normal(mu[0], sigma[0],
nsamp[0]).transpose()
points[:, nsamp[0]:] = np.random.multivariate_normal(mu[1], sigma[1],
nsamp[1]).transpose()
# labels
labels = np.ones((sum(nsamp), 1))
labels[nsamp[0]:] = -labels[nsamp[0]:]
# cost function
z = Variable(dim + 1)
A = np.ones((dim + 1, 1))
A[-1] = 0
obj_func = (C / (2 * NN)) * SquaredNorm(A @ z)
for j in range(sum(nsamp)):
e_j = np.zeros((sum(nsamp), 1))
e_j[j] = 1
A_j = np.vstack((points @ e_j, 1))
obj_func += Logistic(-labels[j] * A_j @ z)
#####################
# Distributed algorithms
#####################
# local agent and problem
EE_min = 1 # kWh
EE_max = rnd(8, 16) # kWh
EE_init = rnd(0.2, 0.5) * EE_max # kWh
EE_ref = rnd(0.55, 0.8) * EE_max # kWh
zeta_u = 1 - rnd(0.015, 0.075) # pure number
#####################
# Generate problem object
#####################
# normalize unit measures
DeltaT = DeltaT / 60 # minutes -> hours
CC_u = CC_u / 1e3 # Euro/MWh -> Euro/KWh
# optimization variables
z = Variable(2 * TT +
1) # stack of e (state of charge) and u (input charging power)
e = np.vstack((np.eye(TT + 1), np.zeros(
(TT, TT + 1)))) @ z # T+1 components (from 0 to T)
u = np.vstack((np.zeros(
(TT + 1, TT)), np.eye(TT))) @ z # T components (from 0 to T-1)
# objective function
obj_func = PP * (CC_u @ u)
# coupling function
coupling_func = PP * u - (PP_max / NN)
# local constraints
e_0 = np.zeros((TT + 1, 1))
e_T = np.zeros((TT + 1, 1))
e_0[0] = 1
in_weights=W[local_rank, :].tolist())
# local variable dimension - random in [2,5]
n_i = np.random.randint(2, 6)
# number of coupling constraints
S = 3
# generate a positive definite matrix
P = np.random.randn(n_i, n_i)
while not is_pos_def(P):
P = np.random.randn(n_i, n_i)
bias = np.random.randn(n_i, 1)
# declare a variable
x = Variable(n_i)
# define the local objective function
fn = QuadraticForm(x - bias, P)
# define the local constraint set
constr = [x>=-2, x<=2]
# define the local contribution to the coupling constraints
A = np.random.randn(S, n_i)
coupling_fn = A.transpose() @ x
# create local problem and assign to agent
pb = ConstraintCoupledProblem(objective_function=fn,
constraints=constr,
coupling_function=coupling_fn)
def generate_LP(n_var: int,
n_constr: int,
radius: float,
direction: str = 'min',
constr_form: str = 'ineq'):
"""Generate a feasible and not unbounded Linear Program and return problem data (cost, constraints, solution)
TODO add reference
Args:
n_var (int): number of optimization variables
n_constr (int): number of constraints
radius (float): size of feasible set (inequality form), size of dual feasible set (equality form)
direction (str): optimization direction - either 'max' (for maximization) or 'min' (for minimization, default)
constr_form (str): form of constraints - either 'ineq' (for inequality: Ax <= b, default) or 'eq' (for standard form: Ax = b, x >= 0)
Returns:
c (np.ndarray): cost vector
A (np.ndarray): constraint matrix
b (np.ndarray): constraint vector (right-hand side)
solution (np.ndarray): optimal solution of problem
"""
size_1 = n_constr
size_2 = n_var
x = Variable(n_var)
# swap dimensions if in dual form (equality constraints)
if constr_form == 'eq':
size_1, size_2 = size_2, size_1
count = 0
while True:
count += 1
# generate a problem in the form min_y c'y s.t. Ay <= b
A_gen = np.random.randn(size_1, size_2)
b_gen = np.random.randn(size_1, 1)
c_gen = A_gen.transpose() @ np.random.randn(size_1, 1)
# prepare problem data
if constr_form == 'ineq':
# primal problem
A = A_gen
b = b_gen
c = c_gen
constr = A_gen.transpose() @ x <= b_gen
else:
# dual problem: min_x b'x s.t. A'x = -c, x >= 0
A = A_gen.transpose()
b = -c_gen
c = b_gen
constr = [A_gen @ x == -c_gen, x >= 0]
# form optimization problem object
if direction == 'min':
obj = c @ x
else:
obj = -c @ x
prob = LinearProblem(objective_function=obj, constraints=constr)
# solve problem to check whether an optimal solution exists
try:
solution = prob.solve()
break
except:
continue
return c, A, b, solution
# get MPI info NN = MPI.COMM_WORLD.Get_size() agent_id = MPI.COMM_WORLD.Get_rank() # Generate a common graph (everyone uses the same seed) Adj = binomial_random_graph(NN, p=0.03, seed=1) ##################### # Problem parameters ##################### np.random.seed(10*agent_id) # linear objective function dim = 2 z = Variable(dim) c = np.ones([dim,1]) obj_func = c @ z # constraints are circles of the form (z-p)^\top (z-p) <= 1 # equivalently z^\top z - 2(A p)^\top z + p^\top A p <= 1 I = np.eye(dim) p = np.random.rand(dim,1) r = 1 # unitary radius constr = [] ff = QuadraticForm(z,I,- 2*(I @ p),(p.transpose() @ I @ p) - r**2) constr.append(ff<= 0) ##################### # Distributed algorithms
# number of columns for each processor
k = 2
# generate a feasible optimization problem of size k * nproc
c_glob, A_glob, b_glob, x_glob = generate_LP(k * nproc, n_constr, 50, constr_form='eq')
# extract the columns assigned to this agent
local_indices = list(np.arange(k*local_rank, k*(local_rank+1)))
c_loc = c_glob[local_indices, :]
A_loc = A_glob[:, local_indices]
b_loc = b_glob
# define the local problem data
x = Variable(k)
obj = c_loc @ x
constr = A_loc.transpose() @ x == b_loc
problem = LinearProblem(objective_function=obj, constraints=constr)
# create agent
agent = Agent(in_neighbors=np.nonzero(Adj[local_rank, :])[0].tolist(),
out_neighbors=np.nonzero(Adj[:, local_rank])[0].tolist())
agent.problem = problem
# instantiate the algorithm
algorithm = DistributedSimplex(agent, enable_log=True, problem_size=nproc*k,
local_indices=local_indices, stop_iterations=2*graph_diam+1)
# run the algorithm
x_sequence, J_sequence = algorithm.run(iterations=100, verbose=True)
np.random.seed()
agent = Agent(in_neighbors=np.nonzero(Adj[local_rank, :])[0].tolist(),
out_neighbors=np.nonzero(Adj[:, local_rank])[0].tolist(),
in_weights=W[local_rank, :].tolist())
# variable dimension
d = 4
# generate a positive definite matrix
P = np.random.randn(d, d)
while not is_pos_def(P):
P = np.random.randn(d, d)
bias = np.random.randn(d, 1)
# declare a variable
x = Variable(d)
# define the local objective function
fun = QuadraticForm(x - bias, P)
# local problem
pb = Problem(fun)
agent.set_problem(pb)
# instantiate the algorithms
initial_condition = np.random.rand(d, 1)
algorithm = GradientTracking(agent=agent,
initial_condition=initial_condition,
enable_log=True)
