Reproduce “Recurrent Environment Simulators” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Silvia Chiappa, Sébastien Racaniere, Daan Wierstra, Shakir Mohamed
DeepMind researchers developed recurrent neural network architectures for simulating how high-dimensional environments respond to actions, achieving spatially and temporally coherent predictions for hundreds of time-steps across diverse tasks like Atari games and 3D mazes. The work also introduced a computationally efficient prediction-independent simulator and demonstrated its utility for model-based exploration.
Key insights
- Increasing the proportion of consecutive prediction-dependent transitions during training generally improves long-term prediction accuracy by focusing model resources on global dynamics, though it can reduce short-term sharpness.
- Direct and global incorporation of action information into the state transition function is more effective than local or indirect methods for maintaining dynamic accuracy.
- Prediction-independent simulators offer significant computational savings by decoupling state evolution from high-dimensional visual frame generation, making long-term planning more efficient for model-based agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.