Sparsely Changing Latent States for Prediction and Planning in Partially Observable Domains
This work addresses the challenge of handling hidden factors in partially observable environments for AI and robotics, offering an incremental improvement in recurrent neural network design.
The paper tackled the problem of prediction and planning in partially observable domains by proposing GateL0RD, a recurrent architecture that enforces sparsely changing latent states, resulting in competitive or superior performance compared to state-of-the-art RNNs in various tasks, with improvements in generalization and sampling efficiency.
A common approach to prediction and planning in partially observable domains is to use recurrent neural networks (RNNs), which ideally develop and maintain a latent memory about hidden, task-relevant factors. We hypothesize that many of these hidden factors in the physical world are constant over time, changing only sparsely. To study this hypothesis, we propose Gated $L_0$ Regularized Dynamics (GateL0RD), a novel recurrent architecture that incorporates the inductive bias to maintain stable, sparsely changing latent states. The bias is implemented by means of a novel internal gating function and a penalty on the $L_0$ norm of latent state changes. We demonstrate that GateL0RD can compete with or outperform state-of-the-art RNNs in a variety of partially observable prediction and control tasks. GateL0RD tends to encode the underlying generative factors of the environment, ignores spurious temporal dependencies, and generalizes better, improving sampling efficiency and overall performance in model-based planning and reinforcement learning tasks. Moreover, we show that the developing latent states can be easily interpreted, which is a step towards better explainability in RNNs.