Generating neural architectures from parameter spaces for multi-agent reinforcement learning

We explore a data-driven approach to generating neural network parameters to determine whether generative models can capture the underlying distribution of a collection of neural network checkpoints. We compile a dataset of checkpoints from neural networks trained within the multi-agent reinforcement learning framework, thus potentially producing previously unseen combinations of neural network parameters. In particular, our generative model is a conditional transformer-based variational autoencoder that, when provided with random noise and a specified performance metric – in our context, returns – predicts the appropriate distribution over the parameter space to achieve the desired performance metric. Our method successfully generates parameters for a specified optimal return without further fine-tuning. We also show that the parameters generated using this approach are more constrained and less variable and, most importantly, perform on par with those trained directly under the multi-agent reinforcement learning framework. We test our method on the neural network architectures commonly employed in the most advanced state-of-the-art algorithms.