Hyperparameters fine tuning for MARL comparative study [D]
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一位研究人员正在使用VMAS库中的多智能体任务训练PPO变体，例如Independent PPO和Graph PPO。他们注意到，在不同的架构/场景组合中，最优超参数（如学习率、熵系数、KL系数、SGD批次大小）往往会有所不同。研究人员质疑，为了公平比较不同架构，是否需要统一所有模型的超参数，并指出统一超参数有时会导致模型不收敛。…
hello everyone. I'm training PPO variants on different multi-agent tasks from the VMAS library (Independent PPO / Graph PPO and such, see HetGPPO by Bettini et al.).
I noticed that for every architecture/scenario couple, the optimal hyperparameters sometimes tend to vary (learning rate, entropy coefficient, KL coefficient, SGD batch size, etc).
do I need - methodologically speaking - to unify the hyperparameters of all models in order to make a fair and correct comparison of architectures later on?
note: sometimes unifying these HP leads to some non converging models.
note 2: my objective is to test these models' robustness under adversarial attack in test-time (frozen models).