Citation Context

...oximation does not occur until a current observation point in need of mapping is present. The simplest Lazy Learning technique is kNN, which is applied to action selection within robotic marble-maze [=-=Bentivegna, 2004-=-] and simulated ball interception [Argall et al., 2007] domains. More complex approaches include Locally Weighted Regression (LWR) [Cleveland and Loader, 1995]. One LWR technique further anchors local...

Citation Context

.... There are three core approaches to policy derivation from demonstration data. In the first approach, the data is used to directly approximate the underlying function mapping observations to actions =-=[5]-=-. In the second approach, the data is used to determine the world dynamics model T (s ′ |s, a) and possibly a function R(s) associating reward with world state [4]. In the third approach, a sequence o...

Citation Context

...behaviors sparsely with via points. Local learning was used for motion control policy development by Atekson et al. [2]. The work presented in this paper also uses local learning techniques, and like =-=[5, 8]-=- combines this with teaching by demonstration. In general, the example executions of a teacher will not apply directly to a learner. Many techniques rely upon an unknown mapping from teacher observati...

Citation Context

... dx, dy) where x and y specify the 2d position on the plane and dx, dy specify the 2d velocity. Actions are also two dimensional (fx, fy) and are force vectors to be applied to the marble. This is not identical but similar to tilting the board. Other simplifications are made in the simulator: the marble is only a point and has no physical extent, the physics are simulated as a sliding block (simplifies friction and inertia) and the simulator adds no artificial noise. Hence, all actions are deterministic. A more realistic but also higher dimensional marble maze simulator was used by Bentivegna [19]. The reward structure used for reinforcement learning in this domain is very simple. Reaching the goal results in a large positive reward. Falling into a hole terminates the trial and results in a large negative reward. Additionally, each action incurs a small negative reward. The agent tries to maximize the reward received, resulting in policies that roughly minimize the time to reach the goal while avoiding holes. Solving the maze from scratch was done using value iteration. In value iteration, dynamic programming sweeps across all states and performs the following update to the value funct...

Citation Context

...ted by observing a sequence of critical events such as the ball just hit a single wall or the ball is against two walls. For more details on our method of segmenting observed data into primitives see =-=[14]-=-. The policies used in the Action Generation module have previously been learned from observing and practicing similar mazes and were also used in the research of [8], [9]. Execution of a primitive en...

Citation Context

... the form of behavior primitives, or simpler policies that contribute to the execution of a more complex policy. Examples include hand-coded primitives used within [9] or automatically extracted from =-=[11]-=- demonstrated tasks, and primitives learned from demonstration [10]. In our work, policy reuse takes a bootstrapping, rather than behavior-primitives, form: the adapted policy performs a different tas...

Citation Context

...te complexity of planning [6], ascertain how to ignore extraneous data in constructing plans from observed task executions [7], [8], or learning policies for executing tasks with performance criteria =-=[9]-=-. Subsequently, research into “good” sets of primitive robot actions has been relatively overlooked; researchers [4], [7] tend to propose new primitives in an ad hoc manner. In contrast, we propose a ...