Schaal, S., & Atkeson, C. (1994). Assessing the quality of learned local models. Advances in Neural Information Processing Systems 6 (pp. 160--167). Morgan-Kaufmann.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....spends approximately one minute generating interaction behaviors. At that phase of the interaction, the goal is to attract as many people as possible for the next tour. Minerva s learning algorithm is a version of reinforcement learning [25] with undelayed reward, using memory based learning [18, 22] for generalization. The state comprises the distance and density of people standing near the robot, as measured by the robot s laser range finder. Actions during interaction, are defined to be joint settings of three features: a facial expression, a pan tilt target for pointing the Feature ....

S. Schaal and C. G. Atkeson. Assessing the quality of learned local models. In J.D. Cowan, G. Tesauro, and J. Alspector, editors, Advances in Neural Information Processing Systems 6, San Mateo, CA, 1994. Morgan Kaufmann.

....used to build global geometrical maps directly. In our method the local grid is used to approximate obstacle boundaries by straight lines. These lines are then used to build the variable resolution partitioning. ffl Real time Learning : We take a memory based learning approach (e.g. 10] 8] [11]) to build the map on line and in real time. A memory based learner is trained by simply storing data in memory, reducing the time needed to incorporate new knowledge in the model. Purely memory based methods do not attempt any data compression (e.g. 8] 11] The proposed method attempts to ....

....learning approach (e.g. 10] 8] 11] to build the map on line and in real time. A memory based learner is trained by simply storing data in memory, reducing the time needed to incorporate new knowledge in the model. Purely memory based methods do not attempt any data compression (e.g. 8] [11]) The proposed method attempts to optimize resources (i.e. memory and time to manage data) by collecting only significant experiences; that is, those experiences that actually improve the model accuracy. ffl Exploration: In order to optimize the learning time, we use an active learning approach ....

S. Schaal and C.G. Atkeson, "Assessing the quality of learned local models," in Neural Information Processing Systems 6, J. Cowan, G. Tesauro, and J. Alspector, Eds., pp. 160--167. Morgan Kaufmann, San Mateo, CA, 1994.

....to minimize the problem of catastrophic interference that would arise in the case of using a global monolithic representation of the environment. ffl Memory based Learning : in order to build the map on line and in real time the robot learns by explicitly memorizing relevant experiences (e.g. [9], 3] A memory based learner is trained by simply storing data, what reduces the time needed to incorporate new knowledge in the current model. Purely memory based methods do not attempt any data compression [9] Our method, on the contrary, optimizes resources (memory and time to manage data) ....

....in real time the robot learns by explicitly memorizing relevant experiences (e.g. 9] 3] A memory based learner is trained by simply storing data, what reduces the time needed to incorporate new knowledge in the current model. Purely memory based methods do not attempt any data compression [9]. Our method, on the contrary, optimizes resources (memory and time to manage data) by collecting significant experiences only [3] that is, experiences that actually improve the model accuracy. ffl Active Learning for Efficient Exploration: an active learning system is one that uses its ....

S. Schaal and C.G. Atkeson. "Assessing the quality of learned local models". In J. Cowan, G. Tesauro, and J. Alspector (eds.), Advances in Neural Information Processing Systems 6 , pp. 160--167, Morgan Kaufmann, 1994.

....to learn sets of polynomial functions, b) incremental learning, and c) support for irrelevant and unscaled features. These requirements directly contribute to the applicability of local learning algorithms (specifically, instance based learning algorithms) Global parametric learning algorithms (Schaal, 1994) such as neural networks attempt to establish an input output mapping via a single function y = f (x, q) where q is a finite length parameter vector. While these methods can theoretically approximate any continuous function (Funahashi, 1989; Hecht Nielson, 1989; Skapura, 1996) they may not be ....

....This is likely to cause problems for learning algorithms that attempt to capture concepts at a global level. Local parametric algorithms attempt to overcome some of the problems of global parametric learning by dividing the input space into many partitions (Atkeson, Schaal, and Moore, 1997; Schaal, 1994). Each partition i is now approximated by an independent function y i = f i (x, q i ) the functions f i are kept as simple as possible. The problem now shifts to the selection of appropriate partitions for the learning system (Schaal and Atkeson, 1994) Nonparametric algorithms (e.g. Atkeson, ....

[Article contains additional citation context not shown here]

Schaal, S., and Atkeson, C. G. 1994. Assessing the Quality of Learned Local Models. In Advances in Neural Information Processing Systems 6, Cowan, J., Tesauro, G., and Alspector J., eds. Morgan Kaufmann.

....most significant induced equations are used as constraints which define the generalized operator s trajectory. Learning the system dynamics The system s dynamics can be learned by any nonlinear function approximator from some execution traces. We use locally weighted regression (Cleveland 1979; Schaal Atkeson 1994), since it enables incremental learning and provides local linear model of the system s dynamics near the current state. Locally weighted regression is a kind of memory based learning, so no generalization is actually done in the training phase. We just need to store the observed points in the ....

Schaal, S., and Atkeson, C. 1994. Assessing the quality of learned local models. In Cowan, J.; Tesauro, G.; and J., A., eds., Advances in Neural Information Processing Systems, volume 6. Morgan Kaufmann.

....example traces. However, experiments show that inducing generalized trajectories leads to much more robust controllers. 2.3 Learning the system s dynamics The system s dynamics can be learned by any nonlinear function approximator from some execution traces. We use locally weighted regression [5, 13], since it enables incremental learning and provides local linear model of the system s dynamics near the current state. Locally weighted regression is a kind of memory based learning, so no generalization is actually done in the training phase. We just need to store the observed points in the ....

S.A. Schaal and C.G. Atkeson. Assessing the quality of learned local models. In J.D. Cowan, G. Tesauro, and Alspector J., editors, Advances in Neural Information Processing Systems, volume 6. Morgan Kaufmann, 1994.

....either plus or minus for example) it is possible to consider more equations, since all of them reflect the relations between the state variables. One such example of combining more equations is presented in the next section. Second the system dynamics is learned using locally weighted regression [12] from some execution traces. Since locally weighted regression is a kind of memory based learning, no learning is actually done in the training phase. We just need to store the observed points in the state space. In the prediction phase, points in the state space are weighted according to the ....

S. Schaal and C.G. Atkeson, "Assessing the quality of learned local models," in Advances in Neural Information Processing Systems 6. 1995, Morgan Kaufmann.

....parametric learning by dividing the input space into many partitions [2, 24] Each partition i is approximated by an independent function y i = f i (x; i ) the functions f i are kept as simple as possible. The problem now shifts to the selection of appropriate partitions for the learning system [25]. Non parametric algorithms (e.g. 2, 6, 22] address this issue by allowing the number of partitions (and consequently the number of parameters) to change dynamically. Instance based learning (IBL) algorithms achieve this by recomputing a fixed set of parameters as a function of the query ....

S. Schaal and C. G. Atkeson. Assessing the quality of learned local models. In J. Cowan, G. Tesauro, and A. J., editors, Advances in Neural Information Processing Systems, volume 6, pages 160--167. Morgan Kaufmann, 1994.

....here use the weighting function 1 1 (x Gammax i ) 2 , but any function that weights the nearer neighbors more will work. The shape of the weighting function also affects the form of the resulting approximation. Examples of robot learning with local linear models can be found in [Atkeson, 1991; Schaal and Atkeson, 1993a; Schaal and Atkeson, 1994] Retrieval efficiency and other limitations Memory based function approximation is attractive because it follows a policy of least commitment. When it receives training data it makes no decision about how to use it for future queries, and just stores the data ....

S. Schaal and C. Atkeson, "Assessing the Quality of Learned Local Models," In Advances in Neural Information Processing Systems (NIPS-6), 1993.

....3 Approximating the operator s value function Our aproach to behavioural cloning in the LQR framework relies on the following observations: 1. An approximate model of the controlled system, in the form of a linear state transition function T , can be induced by locally weighted regression [ Schaal and Atkeson, 94 ] from some available execution traces. 2. Generally, when (a) cost matrices Q and R are given and the execution cost is defined as LQ cost (function of Q and R) and (b) the system is linear, formulas from control theory can be used to compute the state value function V (x) and corresponding ....

S. Schaal and C.G. Atkeson, Assessing the quality of learned local models, Advances in Neural Information Processing Systems 6., Morgan Kaufmann, 1994

....the trajectory the operator is trying to follow is learned from one or more traces of the same subject. We used ML program called GoldHorn (Krizman et al. 1995) to learn the relations between variables in the execution trace. Second the system dynamics is learned using locally weighted regression (Schaal and Atkeson 1995) from some execution traces. Since locally weighted regression is a kind of memory based learning, no learning is actually done in the training phase. We just need to store the observed points in the state space. In the prediction phase, points in the state space are weighted according to the ....

Schaal, S. and C.G. Atkeson (1995). Assessing the quality of learned local models. In: Advances in Neural Information Processing Systems 6.

.... recently that there are significant benefits in terms of data and computational efficiency when data from running the system is used to build a model, rather than using it once for single value function updates (as Q learning would do) and discarding it [ Sutton, 1990, Moore and Atkeson, 1993, Schaal and Atkeson, 1993, Davies, 1996 ] The dynamic programming sweeps can then be done on the learned model either offline or online. In its vanilla form, this method assumes the model is correct and does deterministic dynamic programming using the model. This assumption is often not correct, especially in the early ....

S. Schaal and C. Atkeson. Assessing the quality of learned local models. In Advances in Neural Information Processing Systems (NIPS-6), 1993.

....information about the corresponding region of the world. A local approach also reduces the problem of catastrophic interference that would arise in the case of using a global monolithic representation of the environment. ffl Real time Learning : we take a memory based learning approach (e.g. [6, 15, 21]) to build the environmental map on line and in real time. A memory based learner is trained by simply storing data in memory, what reduces the time needed to incorporate new knowledge in the current model. Purely memory based methods do not attempt any data compression (e.g. 15, 21] Our ....

....(e.g. 6, 15, 21] to build the environmental map on line and in real time. A memory based learner is trained by simply storing data in memory, what reduces the time needed to incorporate new knowledge in the current model. Purely memory based methods do not attempt any data compression (e.g. [15, 21]) Our method, on the contrary, optimizes resources (memory and time to manage data) by collecting only significant experiences; that is, those experiences that do improve the model accuracy. ffl Efficient Exploration: in order to optimize the learning time our robots use an active learning ....

S. Schaal and C.G. Atkeson. Assessing the quality of learned local models. In J. Cowan, G. Tesauro, and J. Alspector, editors, Advances in Neural Information Processing Systems 6, pages 160--167. Morgan Kaufmann, San Mateo, CA, 1994.

No context found.

Schaal, S., & Atkeson, C. (1994). Assessing the quality of learned local models. Advances in Neural Information Processing Systems 6 (pp. 160--167). Morgan-Kaufmann.

No context found.

Schaal, S., & Atkeson, C. (1994). Assessing the quality of learned local models. In J. D. Cowan, G. Tesauro, & J. Alspetor (Eds.). Advances in neural information processing systems, 6 (pp. 160--167). San Mateo, CA: Morgan-Kaufmann.

No context found.

Schaal, S., & Atkeson, C. (1994). Assessing the quality of learned local models. Advances in Neural Information Processing Systems 6 (pp. 160--167). Morgan-Kaufmann.

....order to determine the region of validity of the linear model. Employing the PRESS residual error (Myers, 1990) leave one out cross validation for linear systems can be reformulated such that eliminating data from the data set can be avoided; this formulation also holds for locally linear models (Schaal Atkeson, 1994). Since the PRESS residual error does not exist for PLS, equation (6) approximates it by the sum of PRESS residual errors for every projection k using the residual error, res k i 1, from this stage and the projected data point s k i . The vector s k contains all s k i , and the matrix W is a ....

Schaal, S., & Atkeson, C.G. (1994). Assessing the quality of learned local models. In J. Cowan, G. Tesauro, & J. Alspector (Eds.), Advances in Neural Information Processing Systems 6. San Mateo, CA: Morgan Kaufmann, pp. 160-167.

....distance (Wolberg, 1990) K(d) 1 1 d p (31) can be used to approximate functions like Equation 30 and the quadratic hyperbola kernel 1= h 2 d 2 ) with a well defined value at d = 0. Another smoothing weight function is a Gaussian kernel (Deheuvels, 1977; Wand and Schucany, 1990; Schaal and Atkeson, 1994): K(d) exp i Gammad 2 j (32) This kernel also has infinite extent. A related kernel is the exponential kernel, which has been used in psychological models (Aha and Goldstone, 1992) K(d) exp [ Gamma jdj] 33) These kernels have infinite extent, and can be truncated when they become ....

Schaal, S. and Atkeson, C. G. (1994). Assessing the quality of learned local models. In Cowan et al.

No context found.

Schaal, S., & Atkeson, C. G. (1994b). "Assessing the quality of learned local models." In: Cowan, J. , Tesauro, G., & Alspector, J. (Eds.), Advances in Neural Information Processing Systems 6. Morgan Kaufmann.

No context found.

Schaal, S., & Atkeson, C. G. (1994b). "Assessing the quality of learned local models." In: Cowan, J.

....However, there are many other forms of biases. The chosen input output representation, i.e. which variables are inputs to the learning box and which are the outputs, plays a dominant role in the success of learning. A good choice of the input output representation can make learning trivial (e.g. Schaal et al. 1990; Schaal Sternad, 1992) Geman et al. (1992) demonstrate how a bias on the distance metric in a nonparametric learning system can increase the learning performance significantly. In local averaging (FPNL) the type of averaging function, e.g. constant, linear, quadratic, or cubic, is a bias on ....

Schaal, S., & Atkeson, C. G. (1994b). "Assessing the quality of learned local models." In: Cowan, J. , Tesauro, G., & Alspector, J. (Eds.), Advances in Neural Information Processing Systems 6. Morgan Kaufmann.

.... dof X i w i (y i Gamma x T i fi) 1 Gamma w 2 i x T i (X T X) Gamma1 x i 2 (9) where n = X i w 2 i (10) and dof = X i w 4 i x T i (X T X) Gamma1 x i (11) where the matrix X is already weighted, and dof denotes the local degrees of freedom of the regression [55]. In Figure 5b, the same data as in Figure 5a was fitted by adjusting k to minimize MSE cross at each query point. The outcome is much smoother than that of global cross validation, and also the prediction intervals (a measure of uncertainty described below and indicated by the shaded regions in ....

Stefan Schaal and Christopher G. Atkeson. Assessing the quality of learned local models. In Jack D. Cowan, Gerald Tesauro, and Joshua Alspector, editors, Neural Information Processing Systems Conference, number 6 in Advances In Neural Information Processing Systems, pages 160--167, San Mateo, CA, 1994. Morgan Kaufman.

No context found.

Schaal, S., & Atkeson, C. G. (1994b). "Assessing the quality of learned local models." In: Cowan, J. , Tesauro, G., & Alspector, J. (Eds.), Advances in Neural Information Processing Systems 6. Morgan Kaufmann.

No context found.

S. Schaal and C. G. Atkeson, "Assessing the quality of learned local models," in Advances in Neural Information Processing Systems (NIPS 94), 1994, pp. 160--167.

No context found.

S. Schaal and C. G. Atkeson, "Assessing the quality of learned local models," in Advances in Neural Information Processing Systems (NIPS 94), pp. 160--167, 1994.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC