Organizations: Department of Electronics and Computer Engineering, Thapathali Campus, Institute of Engineering, Tribhuvan University, Kathmandu 44600, Nepal
Dynamic spectrum access must coordinate secondary users under bursty primary-user activity while preserving packet reliability and delay. We present TACAN, a centralized policy that represents each channel as a token containing occupancy history and automatic-modulation-classification entropy; a context token supplies queue class, delay and user identity. A Transformer encoder is warm-started from an occupancy-greedy policy and refined with proximal policy optimization. The frozen policies were trained to maintain a channel assignment in every slot, including when queues were empty. We therefore replay them on held-out trajectories and distinguish standby assignment success from packet-present access and packet delivery. In a 20-channel network with 60 primary devices and 4 secondary users, TACAN achieves 92.53% +/- 0.47 packet-present access success, compared with 89.94% for Greedy and 83.53% for PPO+MLP. Its paired gain over Greedy is 2.59 points (parametric 95% CI 1.89-3.29), with wins in all five seeds; the exact two-sided sign-test value is 0.0625. The gain rises from 0.57 points at normal primary-user load to 7.67 points at extreme load. TACAN also reduces mean delivery delay from 1.208 to 1.123 slots and the conditional user-reliability gap from 9.69 to 3.15 points. Delivered packets per SU-slot remain arrival-limited (30.12% versus 30.11%), so no packet-throughput gain is claimed.
Figures & tables
Work
Att. target
Where
QoS
PHY
SDN
[ 9 , 2 ]
none
n/a
no
no
no
[ 10 , 12 ]
none
n/a
no
no
no
[ 22 , 21 ]
sensing reports
detector
no
I/Q
no
[ 20 ]
sensor nodes
detector
no
I/Q
no
[ 23 ]
users
critic
no
no
no
[ 4 , 5 ]
graph neighbours
policy
traffic
no
no
TABLE I: Position of TACAN relative to the closest prior work. Att. target is what attention operates over; PHY denotes physical-layer features in the observation.
Class
Pri.
λ
α
Dur.
Dominant mod.
URLLC
3
1/8
0.8
1 to 10
BPSK 40%, QPSK 50%
mMTC
1
1/60
0.5
1 to 5
OOK 35%, BPSK 40%
eMBB
2
1/12
1.5
5 to 40
16QAM 50%, 8PSK 15%
TABLE II: IoT traffic class parameters. Durations are in slots.
Method
Assignment (%)
Packet-present access (%)
Del./slot (%)
Mean delay
P95 delay
User gap (pp)
Random (coordinated)
51.65 ± 0.60
50.26 ± 0.85
29.67
3.760
11.850
1.29
Greedy
91.60 ± 0.32
89.94 ± 0.57
30.11
1.208
2.000
9.69
PPO+MLP
86.11 ± 0.77
83.53 ± 0.98
30.08
1.359
3.000
9.75
TACAN
92.93 ± 0.14
92.53 ± 0.47
30.12
1.123
2.000
3.15
Current-slot genie
88.30 ± 0.41
88.13 ± 0.32
30.10
1.203
2.000
1.45
TABLE III: Packet-aware re-evaluation of the frozen policies (mean ± standard deviation over five seeds). Assignment success includes standby assignments; access success is conditioned on a queued packet. Del./slot is delivered packets per SU-slot, and user gap is the max–min conditional-access gap.
TACAN vs.
Metric
Mean
95% CI
Wins
Exact p
Greedy
Access success (pp)
+2.59
[1.89, 3.29]
5/5
0.0625
Greedy
Mean-delay reduction
+0.09
[0.06, 0.11]
5/5
0.0625
Greedy
User-gap reduction (pp)
+6.54
[5.82, 7.27]
5/5
0.0625
PPO+MLP
Access success (pp)
+9.00
[7.94, 10.06]
5/5
0.0625
PPO+MLP
Mean-delay reduction
+0.24
[0.17, 0.30]
5/5
0.0625
PPO+MLP
User-gap reduction (pp)
+6.60
[4.87, 8.33]
5/5
0.0625
TABLE IV: Paired packet-aware differences for TACAN. Confidence intervals are parametric Student- t intervals; the exact sign test is reported because only five trained seeds are available.
Method
Normal
High
Extreme
Greedy
93.28
89.08
81.29
PPO+MLP
87.67
82.41
74.59
TACAN
93.85
92.30
88.96
Current-slot genie
90.34
87.04
82.69
TABLE V: Packet-present access success (%) by primary-user load, averaged over five seeds.
URLLC
mMTC
eMBB
Method
Access (%)
Delay
Access (%)
Delay
Access (%)
Delay
Greedy
89.72
1.209
90.01
1.211
89.99
1.196
PPO+MLP
83.14
1.353
83.43
1.367
84.26
1.339
TACAN
92.56
1.121
92.53
1.124
92.50
1.121
TABLE VI: Packet-present access success and mean delivery delay by traffic class, averaged over five seeds.
Method
Access gap (pp)
Delivery gap (pp)
Delay gap
Access Jain
Greedy
9.69
0.97
0.312
0.9983
PPO+MLP
9.75
0.98
0.350
0.9980
TACAN
3.15
0.98
0.066
0.9998
TABLE VII: Packet-aware inter-user fairness. Gaps are max–min differences averaged over five seeds.
(b) Input-perturbation ablation (same trained policy, feature suppressed at inference)
Unmodified observation
92.93 ± 0.14
7.07
n/a
− AMC-entropy feature
90.04 ± 1.31
9.96
− 2.89
TABLE VIII: Standby-assignment ablation. Group (a) compares complete retrained encoders; the MLP is smaller and lacks TACAN’s auxiliary head, so the row is not an attention-only ablation. Group (b) suppresses an input modality at inference time for the same trained TACAN policy. ΔA is the assignment-success change in percentage points.
Uniform mix (%)
Assignment success (%)
ΔA (pp)
Failure (%)
0 (original)
92.93 ± 0.14
+0.00
7.07
25
92.89 ± 0.17
-0.04
7.11
50
92.72 ± 0.31
-0.21
7.28
75
92.64 ± 0.35
-0.29
7.36
100
92.59 ± 0.32
-0.35
7.41
TABLE IX: Standby-assignment sensitivity to AMC posterior flattening, over 5 matched seeds. The mixing rate is the fraction of a uniform posterior blended into the simulated AMC output.
Component
Median (ms)
95th pct. (ms)
Simulation state extraction
0.041
0.087
Policy inference
4.943
17.097
Rule-dictionary construction
0.033
0.069
Total measured path
5.01
17.27
TABLE X: Simulation-mode controller decision-path latency (1000 cycles, single CPU core). The measurement excludes sensing, AMC inference, OpenFlow transport and installation, and radio channel switching.
School of Computer Science and Technology, Xidian University, China · 2Research and Development, Ericsson, Canada · 3Hainayun IoT Technology Ltd, China
The University of Hong Kong, Hong Kong · 5GIC & 6GIC, Institute for Communication Systems (ICS), University of Surrey, GU2 7XH Guildford, U.K. · Beijing Institute of Technology, Beijing 100081, China +2