Typically, a Reinforcement Learning (RL) algorithm focuses in learning a single deployable policy as the end product. Depending on the initialization methods and seed randomization, learning a single policy could possibly leads to convergence to different local optima across different runs, especially when the algorithm is sensitive to hyper-parameter tuning. Motivated by the capability of Generative Adversarial Networks (GANs) in learning complex data manifold, the adversarial training procedure could be utilized to learn a population of good-performing policies instead. We extend the teacher-student methodology observed in the Knowledge Distillation field in typical deep neural network prediction tasks to RL paradigm. Instead of learning a single compressed student network, an adversarially-trained generative model (hypernetwork) is learned to output network weights of a population of good-performing policy networks, representing a school of apprentices. Our proposed framework, named Teacher-Apprentices RL (TARL), is modular and could be used in conjunction with many existing RL algorithms. We illustrate the performance gain and improved robustness by combining TARL with various types of RL algorithms, including direct policy search Cross-Entropy Method, Q-learning, Actor-Critic, and policy gradient-based methods.@en

Cross-Entropy Method (CEM) is a gradient-free direct policy search method, which has greater stability and is insensitive to hyperparameter tuning. CEM bears similarity to population-based evolutionary methods, but, rather than using a population it uses a distribution over candidate solutions (policies in our case). Usually, a natural exponential family distribution such as multivariate Gaussian is used to parameterize the policy distribution. Using a multivariate Gaussian limits the quality of CEM policies as the search becomes confined to a less representative subspace. We address this drawback by using an adversarially-trained hypernetwork, enabling a richer and complex representation of the policy distribution. To achieve better training stability and faster convergence, we use a multivariate Gaussian CEM policy to guide our adversarial training process. Experiments demonstrate that our approach outperforms state-of-the-art CEM-based methods by 15.8% in terms of rewards while achieving faster convergence. Results also show that our approach is less sensitive to hyper-parameters than other deep-RL methods such as REINFORCE, DDPG and DQN. @en