-
Notifications
You must be signed in to change notification settings - Fork 1
/
uav_flpo.py
295 lines (278 loc) · 13.4 KB
/
uav_flpo.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
class uav_flpo:
def __init__(self,drones,num_stations,blocks=None,
ugv_factor=0.0,fcr=25,distance='euclidean') -> None:
super().__init__()
self.den_drones = drones
self.drones,lb_ub = normalize_drones(drones)
self.lb,self.ub = lb_ub
self.scale = self.ub-self.lb
self.stations=np.repeat([0.5,0.5],num_stations,axis=0)
self.blocks=normalize_blocks(blocks,lb_ub)
self.N_drones= len(drones)
self.N_stations=num_stations
self.stage_horizon=self.N_stations+1
self.gamma_k_length=self.N_stations+1
self.fcr=fcr/self.scale
self.distance=distance
self.ugv_factor=ugv_factor
self.bounds = [(0, 1)]*self.N_stations*2
self.cost_normalizer = 1/(self.N_drones*(self.N_stations**self.N_stations))
return
def return_stagewise_cost(self,params,beta): #params is like stations
d_F=cdist(params,params,self.distance)
if not self.blocks==None:
for block in self.blocks:
d_F=d_F+cdist(params,params,metric=add_block_dist(block[0],block[1]))
d_F=d_F+diag([my_inf]*self.N_stations)
d_delta_to_f=array([my_inf]*self.N_stations).reshape(1,-1)
d_df=np.concatenate((d_F,d_delta_to_f),axis=0)
D_ss=[0]*self.N_drones
for drone_id,drone in enumerate(self.drones):
stage=concatenate((params,array(self.drones[drone_id][1]).reshape(1,-1)),axis=0)
D_s=[0]*(self.stage_horizon+1)
stage_0=array(self.drones[drone_id][0]).reshape(1,-1)
D_s[0]=cdist(stage_0,stage,self.distance)
if not self.blocks==None:
for block in self.blocks:
D_s[0]=D_s[0]+cdist(stage_0,stage,metric=add_block_dist(block[0],block[1]))
#print(D_s[0]-self.drones[drone_id][2]*self.fcr)
D_s[0]=D_s[0]+penalty(D_s[0]-self.drones[drone_id][2]*self.fcr)
#D_s[0]=D_s[0]+my_exp(beta*(D_s[0]-self.drones[drone_id][2]*self.fcr))
#D_s[0][0,-1]=D_s[0][0,-1]*(D_s[0][0,-1]>my_inf)
delta_id= self.N_stations+drone_id
# so far we have taken care of the first distance matrix
d_f_to_delta=cdist(params,array(self.drones[drone_id][1]).reshape(1,-1),self.distance)
if not self.blocks==None:
for block in self.blocks:
d_f_to_delta=d_f_to_delta+cdist(params,array(self.drones[drone_id][1]).reshape(1,-1),metric=add_block_dist(block[0],block[1]))
d_last=np.concatenate((d_f_to_delta,array([0]).reshape(1,-1)),axis=0)
d=np.concatenate((d_df,d_last),axis=1)
d=d+(penalty(d-self.fcr))
#d=d+my_exp(beta*(d-self.fcr))
D_s[1:self.stage_horizon] = [d] * (self.stage_horizon - 1)
d_l=[my_inf]*(self.gamma_k_length-1)
d_l.append(0.0)
D_s[-1]=array(d_l).reshape(-1,1)
D_ss[drone_id]=D_s
self.D_ss=D_ss
return
def calc_associations(self,beta):
p=[]
self.return_stagewise_cost(self.params.reshape(-1,2),beta)
D_ss=self.D_ss
for D_s in D_ss:
K=len(D_s)
D=D_s[::-1]
out_D=[0]*(K+1)
out_D[0]=array([0.0]).reshape(-1,1)
out_p=[0]*(K+1)
out_p[0]=array([1.0]).reshape(-1,1)
out=[0]*(K+1)
out[0]=array([1.0]).reshape(-1,1)
for i in range(1,K+1):
out_D[i]=(D[i-1]+repeat(transpose(out_D[i-1]),D[i-1].shape[0],axis=0))
m=out_D[i].min(axis=1,keepdims=True)
exp_D=exp(multiply(-beta,out_D[i]-m))
out[i]=sum(multiply(exp_D,tile(out[i-1], (1,D[i-1].shape[0])).T),axis=1,keepdims=True)
out_p[i]=divide(multiply(exp_D,out[i-1].T),out[i])
out_D[i]=m
p.append(out_p[::-1][:-1])
self.P_ss=p
return
def free_energy(self,D_s,P_s,beta):
'''
input: D_s: a list of K numpy arrays corrosponding to distances between stages
P_s: a list of K numpy arrays corrosponding to probabilities between stages
output: out_c: K+1 numpy arrays with shape[1]=1, indicating the total cost of nodes
'''
K=len(D_s)
D=D_s[::-1]
P=P_s[::-1]
out_P=[0]*(K+1)
out_C=[0]*(K+1)
out_H=[0]*(K+1)
out_P[0]=array([1.0]).reshape(-1,1)
out_C[0]=array([0.0]).reshape(-1,1)
out_H[0]=array([0.0]).reshape(-1,1)
for i in range(1,K+1):
# assigning P of each node for calculating C in the next i
out_P[i]=(P[i-1]*repeat(transpose(out_P[i-1]),P[i-1].shape[0],axis=0)).sum(axis=1).reshape(-1,1)
out_C[i]=(P[i-1]*(D[i-1]*repeat(transpose(out_P[i-1]),D[i-1].shape[0],axis=0)+repeat(transpose(out_C[i-1]),D[i-1].shape[0],axis=0))).sum(axis=1).reshape(-1,1)
out_H[i]=-(P[i-1]*(my_log(P[i-1])*repeat(transpose(out_P[i-1]),D[i-1].shape[0],axis=0)-repeat(transpose(out_H[i-1]),D[i-1].shape[0],axis=0))).sum(axis=1).reshape(-1,1)
# D-1/beta*H
return (out_C[-1].T).sum() + (-1/beta)*(out_H[-1].T).sum()
def free_energy_Gibbs(self,D_s,beta):
K=len(D_s)
D=D_s[::-1]
out_D=[0]*(K+1)
out_D[0]=array([0.0]).reshape(-1,1)
out=[0]*(K+1)
out[0]=array([1.0]).reshape(-1,1)
for i in range(1,K+1):
out_D[i]=(D[i-1]+repeat(transpose(out_D[i-1]),D[i-1].shape[0],axis=0))
m=out_D[i].min(axis=1,keepdims=True)
exp_D=exp(multiply(-beta,D[i-1]))
out[i]=sum(multiply(exp_D,tile(out[i-1], (1,D[i-1].shape[0])).T),axis=1,keepdims=True)
out_D[i]=m
if isclose(out[-1],0.0).all():
return m.sum()
else:
return (-1/beta*log(out[-1]).sum())
def objective(self,params,beta):
self.return_stagewise_cost(params.reshape(-1,2),beta)
cost=0
for i in range(len(self.D_ss)):
cost+=self.free_energy_Gibbs(self.D_ss[i],beta)
if self.ugv_factor == 0.0:
return self.cost_normalizer*cost
else:
return self.cost_normalizer*cost+self.ugv_factor*linalg.norm(params.reshape(-1,2)-self.stations)
def optimize_D(self,init_guess,beta,method):
# bounds=(np.min([np.min(drone[0]+drone[1]) for drone in drones]),np.max([np.max(drone[0]+drone[1]) for drone in drones]))*len(init_guess)
result = minimize(self.objective, init_guess,args=(beta,),bounds=self.bounds,method=method)
self.params = result.x
self.cost_fun=result.fun
def calc_routs(self):
O=[]
for i in range(self.N_drones):
m=0
o=[]
for p in self.P_ss[i]:
m=argmax(p[m,:])
o.append(m)
o.pop()
O.append(o)
self.routs=O
def train(self,beta_init=1e-6,beta_f=100,alpha=1.5,purturb=0.1,method='powell',verbos=0):
self.Y_s=[]
self.Betas=[]
self.params=ndarray.flatten(self.stations)
beta=beta_init
# self.return_stagewise_cost(self.params.reshape(-1,2))
old_cost=my_inf
while beta <= beta_f:
count=0
self.params=self.params+np.random.normal(0, purturb, self.params.shape)
self.optimize_D(self.params,beta,method=method) #based on P_ss
count+=1
if verbos:
print(f'Beta: {beta:.4e} Cost: {self.cost_fun:0.5e}')
# if abs(self.cost_fun-old_cost) <= 1e-6:
# print("--Optimization Terminated--")
# break
old_cost=self.cost_fun
beta=beta*alpha
self.Y_s.append(self.params.reshape(-1,2))
self.Betas.append(beta)
self.calc_associations(beta)
self.calc_routs()
def print_routs(self):
print("")
for i,o in enumerate(self.routs):
print(f'\nDrone {i+1} --->', end='')
for j in o:
if j<env.N_stations:
print(f'f{j+1} --->', end='')
else:
print(f'[D{i+1}]', end='')
break
def return_total_cost(self):
return self.cost_fun*self.scale/self.cost_normalizer
def return_direct_cost(self):
return np.sum([np.sum((np.array(drone[0])-np.array(drone[1]))**2)**0.5 for drone in self.den_drones])
def plot_routs(self,show_info=True,show_nums=True,save=False,show_ugv=False):
state_locs=self.params.reshape(-1,2).copy()*self.scale+self.lb
den_drones = self.den_drones
drone_locs=array([i[0] for i in den_drones])
dest_locs=array([i[1] for i in den_drones])
plt.scatter(drone_locs[:,0],drone_locs[:,1],color='black',label='UAVs')
for i, loc in enumerate(drone_locs):
if show_info:
plt.text(loc[0], loc[1], 'V'+str(i+1)+f' ({self.drones[i][2]})', ha='center', va='bottom')
plt.scatter(state_locs[:,0],state_locs[:,1],marker='^',label='UGVs')
f_indices = np.argsort(state_locs[:,0])
for i, loc in enumerate(state_locs):
if show_info:
plt.text(loc[0], loc[1], 'F'+str(i+1), ha='center', va='bottom')
plt.scatter(dest_locs[:,0],dest_locs[:,1],marker='*',label='Dest.')
for i, loc in enumerate(dest_locs):
if show_info:
plt.text(loc[0]+1*np.random.rand()*(-1)**round(np.random.rand()), loc[1]+1*np.random.rand(), 'D'+str(i+1), ha='center', va='bottom')
options = ['-', '--', '-.', ':',]
modified_lines = []
colors = []
styles = []
for i, o in enumerate(self.routs):
line_style=np.random.choice(options)
styles.append(line_style)
drone_loc = drone_locs[i]
try:
state_loc = state_locs[o[0]]
except:
state_loc=dest_locs[i]
color=np.random.rand(3)
colors.append(color)
plt.plot([drone_loc[0], state_loc[0]], [drone_loc[1], state_loc[1]], color=color,linewidth=1.5,linestyle=line_style)
dist = np.sqrt(np.sum((drone_loc - state_loc) ** 2))
col='red' if dist > self.drones[i][2]*self.fcr*self.scale else 'green'
if dist >0.0:
if show_nums:
plt.text((drone_loc[0] + state_loc[0]) / 2, (drone_loc[1] + state_loc[1]) / 2, f'{dist:.2f}', color=col, ha='center')
if len(o)>1:
for j in range(len(o) - 1):
try:
loc1 = state_locs[o[j]]
except:
loc1 = dest_locs[i]
try:
loc2 = state_locs[o[j + 1]]
except:
loc2=dest_locs[i]
plt.plot([loc1[0], loc2[0]], [loc1[1], loc2[1]], color=color,linewidth=1.5,linestyle=line_style)
dist = np.sqrt(np.sum((loc1 - loc2) ** 2))
col='red' if dist > self.fcr*self.scale else 'green'
if dist>0.0:
if show_nums:
plt.text((loc1[0] + loc2[0]) / 2, (loc1[1] + loc2[1]) / 2, f'{dist:.2f}', color=col, ha='center')
if not self.blocks==None:
counter=0
den_blocks = denormalize_blocks(self.blocks,(self.lb,self.ub))
for center, radius in den_blocks:
c_x, c_y = center
label='Obst.' if counter==0 else None
circle=patches.Circle(center, radius,edgecolor="white", linewidth=0.05,facecolor='grey', linestyle='-',label=label)
circle.set_zorder(2)
plt.gca().add_patch(circle)
counter+=1
ax = plt.gca()
# Process each drone, but only draw the arcs this time
counter = 0
label=None
drone_id=0
for drone, route in zip(den_drones, self.routs):
path_segments = get_drone_path_segments(drone, route, state_locs)
for segment in path_segments:
for block in den_blocks:
intersections = line_circle_intersection(block[0], block[1], segment[0], segment[1])
if intersections:
if counter==0:
label='Adj.'
counter +=1
theta_start, theta_end = calculate_shorter_arc(block[0], block[1], intersections[0], intersections[1])
arc = patches.Arc(block[0], 2*block[1], 2*block[1],
theta1=np.degrees(theta_start), theta2=np.degrees(theta_end),
color=colors[drone_id],linewidth=3, linestyle=styles[drone_id])
ax.add_patch(arc)
label=None
drone_id +=1
if show_ugv:
plt.scatter(self.stations[0,0],self.stations[0,1],color='orange',marker='2')
plt.text(self.stations[0,0], self.stations[0,1], 'Initial UGV location', ha='center', va='bottom')
plt.xlabel('X')
plt.ylabel('Y')
plt.legend()
plt.title(f'M={self.N_stations} F.C.R. = {self.fcr*self.scale}')
plt.gca().set_aspect('equal')
if save:
plt.savefig(f'sim res for M={self.N_stations} F.C.R. = {self.fcr*self.scale}.pdf')
plt.show()