Nonparametric Representation of Policies and Value Functions: A Trajectory-Based Approach

By Christopher G. Atkeson, Jun Morimoto. 2002

90 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Dual Policy Iteration (2018, methodology): “The instantiation shares similar spirit from some previous works from robotics and control literature, including works from [17, 18] and Guided Policy Search (GPS) [19] (and its variants (e.”
A generalized iterative LQG method for locally-optimal feedback control of constrained nonlinear stochastic systems (2005, background): “DDP was recently applied to the hard problem of robot locomotion, with remarkable success [9].”
Policy Search by Dynamic Programming (2003, methodology): “Samples ofμ are generated in the same way [2] generates initial trajecto ri s, using a parametric policy search.”
An Invitation to Imitation (2015, background): “Atkeson [Atkeson and Morimoto, 2003 ], and others, notably Figure 1 : Human expert demonstration to train a walking robot to cross very rough terrain.”
Space-indexed dynamic programming: learning to follow trajectories (2008, methodology): “Then, they use dynamic programming to compute controllers for timesteps T 1, T 2 and so on down to time-step 1. Some examples of algorithms in this family include (Jacobson & Mayne, 1970; Bagnell et al., 2004; Atkeson & Morimoto, 2003; Lagoudakis & Parr, 2003), and all of them output time-varying/”
Optimal control with learned local models: Application to dexterous manipulation (2016, background): “Previous work has addressed this issue by combining multiple linear-Gaussian controllers into a single nonlinear policy, using methods such as guided policy search [9] and trajectory-based dynamic programming [25].”
Receding Horizon Differential Dynamic Programming (2007, methodology): “In [4] the idea of using the second order local DDP models to make locally-linear controllers is introduced.”
Guided Policy Search (2013, methodology): “While several TBDP methods have been proposed (Atkeson & Morimoto, 2002; Atkeson & Stephens, 2008), we used a simple nearest-neighbor variant with a single trajectory, which is suitable for episodic tasks with a single initial state.”
E-Graphs: Bootstrapping Planning with Experience Graphs (2012, methodology): “In [1], new trajectories in a state space were generated by combining nearby trajectories, including information about local value fun ction estimates at waypoints.”
Model learning for robot control: a survey (2011, methodology): “These approximations are used to improve an open-loop policy where the linearizations are also updated after every policy update (Atkeson and Morimoto 2002).”
Robust Locally-Linear Controllable Embedding (2017, background): “Algorithms from this class such as differential dynamic programming (DDP) [3], iterative linear quadratic regulator (iLQR) [8], and iterative linear quadratic Gaussian (iLQG) [14] have been successfully applied to a variety of complex control problems [1, 13, 7, 9].”
Robust Reinforcement Learning (2005, unknown): “We will apply RRL to more complex tasks like learning stand-up behaviors (Morimoto & Doya, 2000, 2001a) and biped locomotion ( Atkeson & Morimoto, 2003; Morimoto & Atkeson, 2003) in future work.”
Manipulators and Manipulation in high dimensional spaces (2016, unknown): “Researchers have addressed the issue of generalization by combining multiple linearGaussian controllers into a single nonlinear policy, using methods such as guided policy search [53] and trajectory-based dynamic programming [8].”
Manifold Embeddings for Model-Based Reinforcement Learning under Partial Observability (2009, background): “Impressive performance and scalability of local modelbased approaches [1, 2] and global model-free approaches [6, 17] have been achieved by exploiting the locality of dynamics in fully observable state-space representations of challenging real-world problems.”
A robot path planning framework that learns from experience (2012, background): “The most common approaches to solving such tasks is to construct a motion plan using no prior knowledge of the task and the environment; a planning algorithm is only given the robot and environment models along with the start and goal configurations and asked to generate a path.”