In plain words: A separate encoder learns a growing set of visual patterns and codes each game screen sparsely, so the decision network only picks actions. On Atari games, tiny 6-to-18-neuron networks matched or sometimes beat standard deep networks with about 100 times more neurons.
Abstract
Deep reinforcement learning, applied to vision-based problems like Atari games, maps pixels directly to actions; internally, the deep neural network bears the responsibility of both extracting useful information and making decisions based on it. By separating the image processing from decision-making, one could better understand the complexity of each task, as well as potentially find smaller policy representations that are easier for humans to understand and may generalize better. To this end, we propose a new method for learning policies and compact state representations separately but simultaneously for policy approximation in reinforcement learning. State representations are generated by an encoder based on two novel algorithms: Increasing Dictionary Vector Quantization makes the encoder capable of growing its dictionary size over time, to address new observations as they appear in an open-ended online-learning context; Direct Residuals Sparse Coding encodes observations by disregarding reconstruction error minimization, and aiming instead for highest information inclusion. The encoder autonomously selects observations online to train on, in order to maximize code sparsity. As the dictionary size increases, the encoder produces increasingly larger inputs for the neural network: this is addressed by a variation of the Exponential Natural Evolution Strategies algorithm which adapts its probability distribution dimensionality along the run. We test our system on a selection of Atari games using tiny neural networks of only 6 to 18 neurons (depending on the game's controls). These are still capable of achieving results comparable---and occasionally superior---to state-of-the-art techniques which use two orders of magnitude more neurons.
Giuseppe Cuccu, Julian Togelius, Philippe Cudre-Mauroux
arXiv:1806.01363 · cs.LG, cs.AI, cs.NE, stat.ML · submitted Jun 4, 2018 · updated Mar 3, 2019
abstract · pdf · html · Accepted at AAMAS 2019
The staging of components in this paper (compressor/controller), where neuroevolution is only applied to a low-dimensional controller, reminds me of Ha and Schmidhuber's recent paper on world models (which is briefly cited) [1]. They employ a variational autoencoder with ~4.4M parameters, an RNN with ~1.7M parameters, and a final controller with just 1,088 parameters! Though it's recently been shown that neuroevolution can scale to millions of parameters [2], the technique of applying evolution to as few parameters as possible and supplementing with either autoencoders or vector quantization seems to be gaining traction. I hope to apply some of the ideas in this paper to multiple co-evolving agents...
[1]. https://worldmodels.github.io
[2]. https://arxiv.org/abs/1712.06567