Warm Starting Neural Dynamics Models From Few Jacobian Samples
TDec-Init reduced training steps between episodes while matching or beating Sobolev training in low-data benchmarks.
The method initializes neural dynamics models for model-based reinforcement learning using transition Jacobians. It decomposes a tensor of stacked Jacobian samples to recover weights and flexible nonlinearities for a feedforward network. Across several benchmark environments with limited data, the authors report better performance than random initialization and value training. The abstract does not quantify the gains or name the benchmark environments.
ArXiv · AI/CL/LG's note
The method initializes neural dynamics models for model-based reinforcement learning using transition Jacobians. It decomposes a tensor of stacked Jacobian samples to recover weights and flexible nonlinearities for a feedforward network. Across several benchmark environments with limited data, the authors report better performance than random initialization and value training. The abstract does not quantify the gains or name the benchmark environments.
ArXiv · AI/CL/LG's note
score 4