Additional ablation studies on Ant is shown in Figure 1a for a thorough comparison. We show in Figure 1 the visualization of the late stage policy of one seed from Walker2d-v2 environment.

In this section, we provide the proofs of the propositions stated in the main text. However, if an'inconsistent' decoder-encoder pair would be used, an encoder with a perturbed mean In the PCA case, the invariant subspace is explicitly known thanks to the linearity. "autoencoding" requires that realizations generated by the decoder are approximately invariant when The algorithm is shown in Algorithm 1. While SE introduced an'external selection mechanism' to generate adversarial examples, the analysis in this appendix shows that the approach could be viewed as a robust Bayesian We can employ a robust Bayesian approach to define a'pessimistic' bound in the sense of selecting With the given tighter bound the algorithm for SE is shown in Algorithm 2. From Equation 18 we This algorithm can be used for post training an already trained V AE. Figure 6 shows the graphical The algorithm is shown in Algorithm 4. We approximate the required expectations by their Monte C.5 SE-A V AE Figure 7 shows the graphical model describing A V AE-SS model. The algorithm is shown in Algorithm 3. We approximate the required expectations by their Monte In this example Section 3.1, we will assume that both the observations Convolutional architectures are stabilized using BatchNorm between each convolutional layer.

A standard ML model is commonly generated by a single method that specifies aspects such as architecture, initialization, training data and hyperparameters configuration. The presented work introduces a novel methodology allowing to define multiple methods as distinct agents. Agents can collaborate and compete to generate and improve ML models for a given tasks. The proposed methodology is demonstrated with the generation and extension of a dynamic modular multitask ML system solving more than one hundred image classification tasks. Diverse agents can compete to produce the best performing model for a task by reusing the modules introduced to the system by competing agents. The presented work focuses on the study of agents capable of: 1) reusing the modules generated by concurrent agents, 2) activating in parallel multiple modules in a frozen state by connecting them with trainable modules, 3) condition the activation mixture on each data sample by using a trainable router module. We demonstrate that this simple per-sample parallel routing method can boost the quality of the combined solutions by training a fraction of the activated parameters.