We introduce the value iteration network (VIN): a fully differentiable neural network with a `planning module' embedded within. VINs can learn to plan, and are suitable for predicting outcomes that involve planning-based reasoning, such as policies for reinforcement learning. Key to our approach is a novel differentiable approximation of the value-iteration algorithm, which can be represented as a convolutional neural network, and trained end-to-end using standard backpropagation. We evaluate VIN based policies on discrete and continuous path-planning domains, and on a natural-language based search task. We show that by learning an explicit planning computation, VIN policies generalize better to new, unseen domains.

I think it would be of significant interest to a substantial portion of the NIPS audience and I recommend that it be accepted.

The Value Iteration Network (VIN) is an end-to-end differentiable architecture that performs value iteration on a latent MDP for planning in reinforcement learning (RL). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a $100\times 100$ maze -- a task which typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introducing an "adaptive highway loss" that constructs skip connections to improve gradient flow. We evaluate our method on both 2D maze navigation environments and the ViZDoom 3D navigation benchmark. We find that our new method, named Dynamic Transition VIN (DT-VIN), easily scales to 5000 layers and casually solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in RL environments.

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

Value Iteration Networks (VINs) have emerged as a popular method to incorporate planning algorithms within deep reinforcement learning, enabling performance improvements on tasks requiring long-range reasoning and understanding of environment dynamics. This came with several limitations, however: the model is not incentivised in any way to perform meaningful planning computations, the underlying state space is assumed to be discrete, and the Markov decision process (MDP) is assumed fixed and known. We propose eXecuted Latent Value Iteration Networks (XLVINs), which combine recent developments across contrastive self-supervised learning, graph representation learning and neural algorithmic reasoning to alleviate all of the above limitations, successfully deploying VIN-style models on generic environments. XLVINs match the performance of VIN-like models when the underlying MDP is discrete, fixed and known, and provides significant improvements to model-free baselines across three general MDP setups.

We introduce the value iteration network (VIN): a fully differentiable neural network with a planning module' embedded within. VINs can learn to plan, and are suitable for predicting outcomes that involve planning-based reasoning, such as policies for reinforcement learning. Key to our approach is a novel differentiable approximation of the value-iteration algorithm, which can be represented as a convolutional neural network, and trained end-to-end using standard backpropagation. We evaluate VIN based policies on discrete and continuous path-planning domains, and on a natural-language based search task. We show that by learning an explicit planning computation, VIN policies generalize better to new, unseen domains.