def parse(self):
for line in sys.stdin:
task_info = line.split()
if task_info[0] == "NODE_COUNT":
self.no_nodes = int(task_info[1])
continue
if task_info[0] != "NODE":
continue
task_type = TaskType.parse_type(task_info[3])
memory = flops = 0
if task_type == TaskType.COMPUTATION:
flops=int(task_info[4])
elif task_type == TaskType.TRANSFER:
memory=int(task_info[4])
t = Task(task_id=int(task_info[1]),
task_type=task_type,
memory=memory, flops=flops)
children = task_info[2].split(',')
if task_type == TaskType.END:
children = []
children = [int(child) for child in children]
if task_type == TaskType.ROOT:
self.dag = DAG()
self.dag.task = t
self.dag.kids_indices = children
else:
self.dag.insert(t, children, None)
class Parser:
"""
@brief : parses input task graph with the following stucture
NODE <index> <children list> <node type> <node cost> <parallelization overhead>
<node type> : ROOT, COMPUTATION, TRANSFER, END
@return : DAG
"""
def __init__(self):
self.no_nodes = 0
self.dag = None
def parse(self):
for line in sys.stdin:
task_info = line.split()
if task_info[0] == "NODE_COUNT":
self.no_nodes = int(task_info[1])
continue
if task_info[0] != "NODE":
continue
task_type = TaskType.parse_type(task_info[3])
memory = flops = 0
if task_type == TaskType.COMPUTATION:
flops=int(task_info[4])
elif task_type == TaskType.TRANSFER:
memory=int(task_info[4])
t = Task(task_id=int(task_info[1]),
task_type=task_type,
memory=memory, flops=flops)
children = task_info[2].split(',')
if task_type == TaskType.END:
children = []
children = [int(child) for child in children]
if task_type == TaskType.ROOT:
self.dag = DAG()
self.dag.task = t
self.dag.kids_indices = children
else:
self.dag.insert(t, children, None)
def genInit_PI_S(self):
self.dag = DAG(self.operation_count) # directed acyclic graph
sort_TPA = self.adjust_Topological_Prior_Assignment(
) # topological sort
sort_RPA = self.adjust_Random_Prior_Assignment(
) # random priority assignment (RPA)
op_priority_dict = {
opId: {
'TPA': p,
'RPA': sort_RPA.pop()
}
for p, opId in enumerate(sort_TPA)
}
sorted_op_priority_dict = sorted(op_priority_dict.items(),
key=lambda x:
(x[1]['TPA'], x[1]['RPA']))
pi_s = [i[0] for i in sorted_op_priority_dict]
return pi_s
class tabuSearch:
def __init__(self,
idx_layout=1,
idx_job=1,
num_AGV=2,
ro=5,
N_iter_fac=200):
self.idx_layout = idx_layout
self.idx_job = idx_job
self.machine_alloc, self.processing_time, self.layout, self.num_AGV = getBenchmarkInstance(
idx_layout, idx_job,
num_AGV) # import instance of JSSPMH (Ulusoy 1995)
self.jobNum = len(self.machine_alloc)
self.machNum = len(self.layout) - 1
self.operation_count = len(np.concatenate(self.machine_alloc, axis=0))
self.operation_id_matrix = self.genOperation_id_Matrix()
self.K = self.operation_count # Tabu list size
self.ro = ro # penalty factor
self.Niter = N_iter_fac * self.operation_count # maximum number of iterations
self.tabu_list = tabuList(self.K) # tabu list
# return operation id 2darray
# operation id is (i,j): job i's jth operation.
def genOperation_id_Matrix(self):
operation_id_matrix = []
for job in range(self.jobNum):
job_list = []
for op in range(len(self.machine_alloc[job])):
job_list.append((job, op))
operation_id_matrix.append(job_list)
return operation_id_matrix
# generate initial feasible pi(solution)
def genInit_PI(self):
pi_s = self.genInit_PI_S()
pi_v = self.genInit_PI_V(pi_s)
pi_init = [pi_s, pi_v]
return pi_init
# generate initial feasible operation pi
# To create a feasible pi_s, topological sort and random priority assignment (RPA) methods are adopted
def genInit_PI_S(self):
self.dag = DAG(self.operation_count) # directed acyclic graph
sort_TPA = self.adjust_Topological_Prior_Assignment(
) # topological sort
sort_RPA = self.adjust_Random_Prior_Assignment(
) # random priority assignment (RPA)
op_priority_dict = {
opId: {
'TPA': p,
'RPA': sort_RPA.pop()
}
for p, opId in enumerate(sort_TPA)
}
sorted_op_priority_dict = sorted(op_priority_dict.items(),
key=lambda x:
(x[1]['TPA'], x[1]['RPA']))
pi_s = [i[0] for i in sorted_op_priority_dict]
return pi_s
# generate initial feasible agv pi
# AGV assignment conforms to the nearest vehicle rule.
def genInit_PI_V(self, pi_s):
pi_v = []
agv_loc_dict = {i: 0 for i in range(self.num_AGV)}
for i in range(self.num_AGV):
pi_v.append(i)
agv_loc_dict[i] = self.machine_alloc[pi_s[i][0]][pi_s[i][1]]
for idx, op in enumerate(pi_s[2:]):
last_op = True if op[1] == len(
self.machine_alloc[op[0]]) - 1 else False
if last_op:
pi_v.append(-1)
else:
mach = self.machine_alloc[op[0]][op[1]]
agv_distance_list = {}
for agv_id, agv_loc in agv_loc_dict.items():
dist = self.layout[agv_loc][mach]
agv_distance_list[agv_id] = dist
if agv_distance_list[0] > agv_distance_list[1]:
near_agv = 1
else:
near_agv = 0
pi_v.append(near_agv)
agv_loc_dict[near_agv] = mach
return pi_v
# topological sort
def adjust_Topological_Prior_Assignment(self):
for each_job in self.operation_id_matrix:
for idx, j in enumerate(each_job):
if idx == len(each_job) - 1:
self.dag.graph[j] = []
else:
self.dag.addEdge(j, each_job[idx + 1])
return self.dag.topologicalSort()
# random priority assignment (RPA)
def adjust_Random_Prior_Assignment(self):
priority_list = [i for i in range(self.operation_count)]
np.random.shuffle(priority_list)
return priority_list
# operation sequence-fixed neighbourhood generation (OSF) -> Changing pi_v
def generate_OSF(self, pi):
pi_s = pi[0]
pi_v = pi[1]
pi_neighbors_list = []
for idx, val in enumerate(pi_v):
neighbor_pi_v = copy.deepcopy(pi_v)
"""
# AGV num >= 3 Case:
other_AGVs = [i for i in range(self.num_AGV) if i!=val]
random_select_AGV = np.random.choice(other_AGVs)
self.pi_v[idx] = random_select_AGV # reallocate
"""
if val != -1:
if val == 0:
neighbor_pi_v[idx] = 1
else:
neighbor_pi_v[idx] = 0
pi_neighbors_list.append([pi_s, neighbor_pi_v])
return pi_neighbors_list
# vehicle assignment-fixed neighbourhood generation (VAF) -> Changing pi_s
def generate_VAF(self, pi):
pi_s = pi[0]
pi_v = pi[1]
pi_neighbors_list = []
# insertion method
for h, val in enumerate(pi_s):
new_h = np.random.choice([i for i in range(len(pi_s)) if i != h])
neighbor_pi_s = copy.deepcopy(pi_s)
neighbor_pi_v = copy.deepcopy(pi_v)
opId = pi_s[h]
vId = pi_v[h]
del neighbor_pi_s[h]
del neighbor_pi_v[h]
neighbor_pi_s.insert(new_h, opId)
neighbor_pi_v.insert(new_h, vId)
neighbor_pi_s, neighbor_pi_v = self.takeRepairMechanism(
h, new_h, opId, neighbor_pi_s, neighbor_pi_v)
neighbor_pi = [neighbor_pi_s, neighbor_pi_v]
pi_neighbors_list.append(neighbor_pi)
return pi_neighbors_list
def takeRepairMechanism(self, h, new_h, opId, pi_s, pi_v):
# Step 1. Among the operations which are located between the original position and
# new position of the inserted operation i, find all operations belonging to J(i).
same_job_op_list = [i for i in pi_s[h:new_h] if i[0] == opId[0]]
# Step 2. Randomly find a valid position for each of those operations of J(i) (found in Step1)
# and insert them into the new positions one by one.
# TODO: update also pi_v
if same_job_op_list:
for op in same_job_op_list:
v = pi_v[pi_s.index(op)]
checked_pos_list = []
while self.checkInfeasible(pi_s, op):
h = pi_s.index(op)
checked_pos_list.append(h)
random_pos = np.random.choice([
idx for idx, val in enumerate(pi_s)
if idx not in checked_pos_list
])
checked_pos_list.append(random_pos)
del pi_s[h]
del pi_v[h]
pi_s.insert(random_pos, op)
pi_v.insert(random_pos, v)
return pi_s, pi_v
def checkInfeasible(self, pi_s, op):
# check precedence constraint of opeartion(op) at pi_s
job_op_top_sort_list = [
i for i in self.dag.topologicalSort() if i[0] == op[0]
]
job_op_pi_s_sort_list = [i for i in pi_s if i[0] == op[0]]
idx_top = job_op_top_sort_list.index(op)
#idx_pi = job_op_pi_s_sort_list.index(job_op_top_sort_list[0])
idx_pi = job_op_pi_s_sort_list.index(op)
job_op_pi_s_sort_filtered_list = [
i for i in job_op_pi_s_sort_list[:idx_pi + 1] if i[1] <= op[1]
]
if job_op_top_sort_list[:idx_top +
1] != job_op_pi_s_sort_filtered_list:
return True
return False
# run the tabu search
def solve(self):
# step 1.
#lmf initialize
lmf = {v: {} for v in range(self.num_AGV)}
for v in range(self.num_AGV):
for job in self.operation_id_matrix:
for op in job:
jobId = op[0]
if op[1] != len(self.machine_alloc[jobId]) - 1:
if op not in lmf[v].keys():
lmf[v][op] = 0
# step 2. create an initial feasible solution. calculate its makespan.
pi_init = self.genInit_PI()
z_init = self.get_objective(pi_init)
pi_curr = pi_init
pi_best = pi_curr
z_curr = z_init
z_best = z_curr
n_iter = 0
# step 3: generate the neighbour solutions
while n_iter < self.Niter:
if n_iter % 20 == 0:
print("iter={}".format(n_iter))
# 3.1 generate neighbour solutions of pi_curr by OSF.
neighbours_OSF_dict = {
idx: {
'sol': sol,
'z': 0
}
for idx, sol in enumerate(self.generate_OSF(pi_curr))
}
# 3.1.1 calculate z_neigh of each neighbour solution.
for idx, val in neighbours_OSF_dict.items():
pi_neigh = val['sol']
z_neigh = self.get_objective(pi_neigh)
neighbours_OSF_dict[idx]['z'] = z_neigh
# For neighbours with z_neigh > z_curr, set z_neigh = z_neigh + ro * lmf(v,i)
for idx, val in neighbours_OSF_dict.items():
pi_neigh = val['sol']
z_neigh = val['z']
if z_neigh > z_curr:
move_idx = [
idx
for idx, v in enumerate(zip(pi_curr[1], pi_neigh[1]))
if v[0] != v[1]
][0]
v = pi_neigh[1][move_idx] # vehicle
i = pi_neigh[0][move_idx] # operation
z_neigh = z_neigh + self.ro * lmf[v][i]
val['z'] = z_neigh
# 3.1.2 find pi_neigh*.
pi_neigh_star = sorted(neighbours_OSF_dict.items(),
key=lambda x: x[1]['z'])[0][1]['sol']
z_neigh_star = self.get_objective(pi_neigh_star)
move_idx = [
idx for idx, v in enumerate(zip(pi_curr[1], pi_neigh_star[1]))
if v[0] != v[1]
][0]
v_star = pi_neigh_star[1][move_idx]
i_star = pi_neigh_star[0][move_idx]
# update the LMF
lmf[v_star][i_star] = lmf[v_star][i_star] + 1
pi_curr = pi_neigh_star
z_curr = z_neigh_star
if z_curr < z_best:
pi_best = pi_curr
z_best = z_curr
# 3.2 generate neighbour solutions of pi_curr by VAF.
# 3.2.1 calculate z_neighbour of each neighbour solution
neighbours_VAF_dict = {
idx: {
'sol': sol,
'z': 0
}
for idx, sol in enumerate(self.generate_VAF(pi_curr))
}
for idx, val in neighbours_VAF_dict.items():
pi_neigh = val['sol']
z_neigh = self.get_objective(pi_neigh)
neighbours_VAF_dict[idx]['z'] = z_neigh
# Neighbours which are not tabu solutions and tabu neighbours with
# objective function value Zneigh < Zbest consist of the admissible neighbour set.
neighbours_admissible_list = []
for idx, val in neighbours_VAF_dict.items():
pi_neigh = val['sol']
z_neigh = val['z']
pi_s = pi_neigh[0]
if pi_s not in self.tabu_list:
neighbours_admissible_list.append([idx, pi_neigh, z_neigh])
if pi_s in self.tabu_list and z_neigh < z_best:
neighbours_admissible_list.append([idx, pi_neigh, z_neigh])
# 3.2.2 find pi_neigh_star in the admissible neighbour set.
_, pi_neigh_star, z_neigh_star = sorted(neighbours_admissible_list,
key=lambda x: x[2])[0]
pi_curr = pi_neigh_star
z_curr = z_neigh_star
if z_curr < z_best:
pi_best = pi_curr
z_best = z_curr
pi_s = pi_neigh_star[0]
self.tabu_list.append(pi_s)
n_iter = n_iter + 1
print("solved makespan is {}".format(z_best))
return pi_best, z_best
# return the objective(makespan) of schedule.
def get_objective(self, pi):
machine_schedule, agv_schedule = self.get_schedule(pi)
makespan = max([i[-1][-1] for i in machine_schedule.values()])
return makespan
# return the schedule of each resource(Machines, AGVs)
def get_schedule(self, pi):
pi_s = pi[0]
pi_v = pi[1]
job_info = {
i: {
'p': p,
'm': m,
'pf': 9999,
'lf': 9999
}
for i, p, m in zip(sorted(pi_s),
np.concatenate(self.processing_time, axis=0),
np.concatenate(self.machine_alloc, axis=0))
}
machine_schedule = {i: [] for i in range(1, len(self.layout))}
agv_schedule = {i: [] for i in range(self.num_AGV)}
# decode
for op, agv in zip(pi_s, pi_v):
mach = job_info[op]['m']
first_op = True if op[1] == 0 else False
last_op = True if agv == -1 else False
if first_op:
job_ready_t = 0
else:
job_ready_t = job_info[op[0], op[1] - 1]['lf']
if not machine_schedule[mach]:
mach_ready_t = 0
else:
mach_ready_t = machine_schedule[mach][-1][-1]
# schedule: Process, Empty Trip, Load Trip
process_st = max(job_ready_t, mach_ready_t)
process_et = process_st + job_info[op]['p']
job_info[op]['pf'] = process_et
if not last_op:
if not agv_schedule[agv]: # agv location: LU stocker
agv_loc = mach
emptyTrip_st = 0
else:
agv_loc = agv_schedule[agv][-1][2]
emptyTrip_st = agv_schedule[agv][-1][-1]
emptyTrip_et = emptyTrip_st + self.layout[agv_loc][mach]
# schedule load trip task
next_mach = job_info[op[0], op[1] + 1]['m']
move_st = max(process_et, emptyTrip_et)
move_et = move_st + self.layout[mach][next_mach]
job_info[op]['lf'] = move_et
agv_schedule[agv].append([(-1, -1), agv_loc, mach,
emptyTrip_st, emptyTrip_et])
agv_schedule[agv].append(
[op, mach, next_mach, move_st, move_et])
machine_schedule[mach].append([op, process_st, process_et])
return machine_schedule, agv_schedule
# return and save the gantt chart of schedule
def draw_Gantt(self, z, machine_schedule, agv_schedule):
import matplotlib.colors as mcolors
color_dict = np.random.choice(list(mcolors.CSS4_COLORS.keys()),
self.jobNum)
job_color = {i: color_dict[i] for i in range(self.jobNum)}
job_color[-1] = 'lightgrey'
fig, gnt = plt.subplots(figsize=(30, 10))
gnt.set_title(
'gantt chart, layout={}, jobset={}, num_AGV={}, makespan={}'.
format(self.idx_layout, self.idx_job, self.num_AGV, z))
gnt.set_xlabel('time')
gnt.set_ylabel('resources')
gnt.set_xticks([i for i in range(100)])
plt.xticks(rotation=45)
gnt.set_yticks([25, 45, 65, 85, 105, 125])
gnt.set_yticklabels(['agv1', 'agv2', 'm1', 'm2', 'm3', 'm4'])
gnt.grid(True)
for k, v in machine_schedule.items():
gnt.broken_barh([(i[1], i[2] - i[1]) for i in v], (k * 20 + 40, 9),
facecolors=[job_color[i[0][0]] for i in v])
for k, v in agv_schedule.items():
gnt.broken_barh([(i[3], i[4] - i[3]) for i in v],
((k + 1) * 20, 9),
facecolors=[job_color[i[0][0]] for i in v])
for i in v:
if i[1] == i[2]:
gnt.text(x=(i[3] + i[4]) / 2, y=k * 20 + 25, s='*wait')
else:
gnt.text(x=(i[3] + i[4]) / 2,
y=k * 20 + 25,
s='{}->{}'.format(i[1], i[2]))
patch_list = []
for i in range(self.jobNum):
patch = mpatches.Patch(color=job_color[i],
label='job_' + str(i + 1))
patch_list.append(patch)
patch = mpatches.Patch(color=job_color[-1], label='empty trip')
patch_list.append(patch)
gnt.legend(handles=patch_list)
fig.savefig('./gantt.png')
# check wheter last operation of each job is unassigned with AGV(=-1)
def checkLastOpFeasible(self, pi):
pi_s = pi[0]
pi_v = pi[1]
last_op_set = [i[-1] for i in self.operation_id_matrix]
for op in last_op_set:
if pi_v[pi_s.index(op)] != -1:
print("infeasible assn, op: {}".format(op))
print("checked")
