Table 1.

Comparison with existing works.

Existing Works	Main Method	Device Mobility	Data Privacy Protection	Network Training Optimization	Optimization Goal	Complexity
[12]	Heuristic, MIP	Static	No	—	Energy consumption and latency	O(N3)
[13]	Game theory	Dynamic	No	—	Energy consumption and latency	O(2N)
[8]	Genetic algorithm, multiactor DQN with single type of agent	Dynamic	No	Original DQN training	Energy consumption and system utility	O(N)
[9]	MINLP, DDPG with single type of agent	Dynamic	No	Original DQN training	System energy consumption	O(TH)
[29]	Q-learning with single type of agent	Static	No	Original Q-learning training with device-to-device communication	Energy consumption and latency	O(TH)
[26]	MINLP, DDQN with single type of agent	Dynamic	No	Original DDQN training	System energy consumption	O(TH)
[27]	Collaboration of FL, blockchain, and DDQL with single type agent	Dynamic	Yes	Decentralized FL training	System latency	O(TH)
Ours	Collaboration of FL and two types of DDQN agent and D3QN agent	Dynamic	Yes	Decentralized FL training with semi-global aggregation	System latency	O(TH)