Fast Value Tracking for Deep Reinforcement Learning

Reinforcement learning (RL) tackles sequential decision-making problems by creating agents that interacts with their environment. However, existing algorithms often view these problem as static, focusing on point estimates for model parameters to maximize expected rewards, neglecting the stochastic dynamics of agent-environment interactions and the critical role of uncertainty quantification. Our research leverages the Kalman filtering paradigm to introduce a novel and scalable sampling algorithm called Langevinized Kalman Temporal-Difference (LKTD) for deep reinforcement learning. This algorithm, grounded in Stochastic Gradient Markov Chain Monte Carlo (SGMCMC), efficiently draws samples from the posterior distribution of deep neural network parameters. Under mild conditions, we prove that the posterior samples generated by the LKTD algorithm converge to a stationary distribution. This convergence not only enables us to quantify uncertainties associated with the value function and model parameters but also allows us to monitor these uncertainties during policy updates throughout the training phase. The LKTD algorithm paves the way for more robust and adaptable reinforcement learning approaches.

Perspective Transformation Layer

Incorporating geometric transformations that reflect the relative position changes between an observer and an object into computer vision and deep learning models has attracted much attention in recent years. However, the existing proposals mainly focus on affine transformations that cannot fully show viewpoint changes. Furthermore, current solutions often apply a neural network module to learn a single transformation matrix, which ignores the possibility for various viewpoints and creates extra to-be-trained module parameters. In this paper, a layer (PT layer) is proposed to learn the perspective transformations that not only model the geometries in affine transformation but also reflect the viewpoint changes. In addition, being able to be directly trained with gradient descent like traditional layers such as convolutional layers, a single proposed PT layer can learn an adjustable number of multiple viewpoints without training extra module parameters. The experiments and evaluations confirm the superiority of the proposed PT layer.