Value Iteration Networks With Gated Summarization Module

In this paper, we address the challenges faced by Value Iteration Networks (VIN) in handling larger input maps and mitigating impact of accumulated errors caused increased iterations. We propose a novel approach, with Gated Summarization Module (GS-VIN), which incorporates two main improvements: 1) employing an Adaptive Strategy module to reduce number iterations; 2) introducing summarize iterative process. The adaptive iteration strategy uses convolution kernels fewer times, reducing network depth increasing training stability while maintaining accuracy planning gated summarization enables emphasize entire process, rather than solely relying on final global outcome, temporally spatially resampling process within VI module. conduct experiments 2D grid world path-finding problems Atari Mr. Pac-man environment, demonstrating that GS-VIN outperforms baseline terms single-step accuracy, success rate, overall performance across different map sizes. Additionally, provide analysis relationship between size, kernel iterations VI-based models, is applicable majority models offers valuable insights for researchers industrial deployment.