在需要训练1e6步的任务中,我一般选择 宽度128、256,层数小于8的网络(请注意,乘以一个w算一层,一层LSTM等于2层,详见曾伊言:LSTM入门例子)。使用ResNet等结构会有很小的提升。一般选择一个略微冗余的网络容量即可,把调整超参数的精力用在这上面不划算,我建议这些超参数都粗略地选择2的N次方,因为: 防止过度调参...
deep-reinforcement-learningdqnmdpoffloadingdeep-q-networkmarkov-decision-processesresource-managementperformance-evaluationmecddqnedge-computinglstm-networksnetwork-optimizationd3qn UpdatedJan 3, 2025 Python OpenAI LunarLander-v2 DeepRL-based solutions (DQN, DuelingDQN, D3QN) ...
deep-reinforcement-learningdqnmdpdeep-q-networkmarkov-decision-processesresource-managementperformance-evaluationmecddqnedge-computinglstm-networksnetwork-optimizationcomputation-offloadingnetwork-performanced3qnqoe-measurementsdouble-deep-q-networkmobile-edge-computingdueling-ddqnqeco ...