Information-theoretic analysis ofdeep neural networks (DNN) has attracted recent interest due to intriguing fundamental results and new hypotheses. Applying information theory to DNNs may provide novel tools for explainable AI via estimation of information flows [1-5], as well as new ways to encourage models to extract and generalize information [1, 6-8].

As aresult, full MOBAgames without restrictions are farfrom being mastered by any existing AI system. In this paper, we propose a MOBA AIlearning paradigm that methodologically enables playing full MOBAgames withdeepreinforcementlearning.Specifically,wedevelopacombinationofnovel and existing learning techniques, including curriculum self-play learning, policy distillation, off-policy adaption, multi-head value estimation, and Monte-Carlo tree-search, intraining andplaying alargepoolofheroes,meanwhile addressing thescalabilityissueskillfully.

Beyond existing meta-RL analyses, we provide upper bounds of the expected optimality gap over the task distribution.

Weuse the twohighest frequencyones which result in 776 label categories. Thelearningrateis12 decreased by afactor of 10 atthe end of 10th and 20th epochs. The networks are trained for 36epochs. Since the all the labels for the test images are not annotated, we only evaluate the performance of17 our model on the set of annotated labels. Hence false positive can happen only if apositively18 annotated label is predicted as a negative class.