L. Meeden. Towards planning: Incremental investigations into adaptive robot control. PhD thesis, Indiana University, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....true gradient descent, but it has been found to work quite well in practice. This has allowed researchers to work with feedback networks without the trouble of the more complex, and more time consuming, learning algorithms. SRNs have been applied to grammar learning (Elman, 1990) robot control (Meeden, 1994), and many other problems requiring recurrent processing. SRNs are related to more elaborate connectionist networks, such as Pollack s Recursive Auto Associative Memory (RAAM; Pollack, 1988) see also Blank, Meeden, and Marshall, 1992, for a thorough examination of sequential RAAMs, a close cousin ....

Meeden, L. (1994). Towards Planning: Incremental Investigations into Adaptive Robot Control. Unpublished Ph.D. dissertation, Indiana University, Department of Computer Science.

....[127] Many researchers have been focusing on such explanations by attempting to solve high level problems via a purely connectionist framework. Some high level systems that come to mind include Elman s developmental models, Meeden s planning system, and my own connectionist analogy making system [43, 133, 20]. Rather than focusing on some assumed necessary symbolically based process (say, variable binding) these models look at a bigger goal: modeling a complex behavior. My analogical model does not assume that analogies are correspondences made via searching through symbolic structures. Rather, a ....

L. Meeden. Towards planning: incremental investigations into adaptive robot control. PhD thesis, Indiana University, Bloomington, 1994. http://www.cs.swarthmore.edu/~ meeden/.

....and perception. This methodology has been explored by a number of robotics researchers (e.g. Aloimonos, 1992; Brooks, 1991b; Coombs, et al. 1995; Horswill, 1993; Pfeifer and Verschure, 1993; Sandini et al. 1993) and has even produced higher order behaviors like planning (e.g. Mataric, 1992; Meeden, 1994). The similarities between these approaches and ours based independently on Gibsonian ideas suggests that the application of the theories and results from fortyfive years of ecological psychology will surely enhance this endeavor. Our work also ties in with recent physiological studies and lesion ....

Meeden, L.A. (1994). Towards planning: Incremental investigations into adaptive robot control. Doctoral Dissertation. Bloomington: Indiana University.

.... are tailored for specific environments of rhythmic patterns, much in the same way that genetic algorithms have been used to evolve connectionist networks to solve specific tasks, such as in learning the weights of a network that controls the movement of a small robot towards a fixed light source (Meeden, 1994). The concept of a genetic algorithm is based on the principles of evolution, operating on a population of individuals, in which each individual represents a suggested solution to a given problem. To solve that problem based only on the randomly generated suggested solutions, the genetic algorithm ....

....of the population, and thus to obtain an improved set of candidate solutions. In order to apply genetic algorithms to the parameterization problem for adaptive oscillators, each adaptive oscillator could be coded as a real valued vector, analogous to the vector coding of network weights used by Meeden (1994). Fitness would be measured by running the coded adaptive oscillator on a test set of rhythmic patterns, selected from the particular environment, and then by averaging the output measure of synchrony for those patterns. As advocated by Meeden (1994) the next generation of individuals would be ....

[Article contains additional citation context not shown here]

Meeden, L. (1994). Toward planning: Incremental investigations into adaptive robot control. Ph.D. Thesis, Indiana University.

....by the system. However, it is difficult to see how Pengi would achieve a longer range goal, requiring coordination of many actions; for instance, in the case of Pengo, the goal of the game requires moving special objects from positions scattered across the board into a particular configuration. Meeden (1994) takes a very different approach to planning and action in Carbot. Meeden trains a recurrent connectionist network using reward and punishment reinforcement learning to control a robot in order to achieve simple goals. For instance, Carbot must learn to alternatively approach and move away from a ....

Meeden, L. (1994). Towards Planning: Incremental Investigations into Adaptive Robot Control. Ph.D. thesis, Indiana University, Computer Science Department.

....between action and perception. This methodology has been explored by a number of robotics researchers (Aloimonos, 1992; Brooks, 1991a; Coombs et al. 1995; Horswill, 1993; Pfeifer Verschure, 1993; Sandini et al. 1993) and has even produced higher order behaviors like planning (Mataric, 1992; Meeden, 1994). The similarities between these approaches and ours based independently on Gibsonian ideas suggests that the application of the theories and results from fifty years of ecological psychology will surely enhance this endeavor. Our work also ties in with recent physiological studies and lesion ....

Meeden, L. A. (1994). Towards planning: Incremental investigations into adaptive robot control.

No context found.

L. Meeden. Towards planning: Incremental investigations into adaptive robot control. PhD thesis, Indiana University, 1994.

....developed within the self organizing map of Level 1. Level 2 is trained to predict what the next Level 1 state will be given its current state. This prediction task will force Level 2 to use its recurrent connections to recognize sequences of sensor motor associations through time. Previous work (Meeden 1994; 1996) has shown that this type of simple recurrent network will develop representations of multi step behaviors, that have been termed protoplans. Level 3 observes the protoplans developed within Level 2 and begins to categorize them. The behavior concepts formed in Level 3 are fed back into ....

Meeden, L. 1994. Towards planning: Incremental investigations into adaptive robot control. Ph.D. Dissertation, Indiana University.

....process. The same intrinsic developmental algorithm can be employed recursively in subsequent stages, using the knowledge discovered in previous stages. For example, the second stage abstracts sequences of behaviors and corresponding perceptual views. These behavior sequences, termed protoplans [6], can lead the robot through a series of views in the environment thus resulting in interesting places to visit. We will call these places protogoals. Here, the proto pre x implies a distinction between standard notions of plans and goals from the developmental ones used here. Once again, a ....

....foundational goal based mechanism, and incrementally bootstrap up toward full scale goals. 7 The level at which purposeful behaviors can be incorporated in our architecture is an open question and remains to be explored. However, we have evidence that protoplans can be used to achieve protogoals [1, 6]. The more interesting issue here is that of goal creation. That is, how does a robot commit itself to achieving a speci c goal We expect the motivational component of the control scheme to play a central role in this. 5 Conclusion The question How does a thing become conscious could be put ....

L. Meeden. Towards planning: Incremental investigations into adaptive robot control. PhD thesis, Indiana University, 1994. 9

....developed within the self organizing map of Level 1. Level 2 is trained to predict what the next Level 1 state will be given its current state. This prediction task will force Level 2 to use its recurrent connections to recognize sequences of sensor motor associations through time. Previous work [16, 17] has shown that this type of simple recurrent network will develop representations of multi step behaviors, that have been termed protoplans. Level 3 observes the protoplans developed within Level 2 and begins to categorize them. The behavior concepts formed in Level 3 are fed back into Level 1 ....

L. Meeden. Towards planning: Incremental investigations into adaptive robot control. PhD thesis, Indiana University, 1994.

....which served as a testbed for the kinds of equipment that have been requested. The PI s research interests include using machine learning techniques such as genetic algorithms, neural networks, and reinforcement learning to adapt the control programs of physical robots. The PI s dissertation [11] described a series of learning experiments with a robot called Carbot aimed to better understand how planning abilities could develop from successful reactive behaviors. The PI is co chairing the international workshop ROBOLEARN96 [5] in association with FLAIRS 96, on learning for autonomous ....

Lisa A. Meeden. Towards planning: Incremental investigations into adaptive robot control. PhD thesis, Indiana University, 1994.

....if it is provided with explicit, abstract goals as input and is adapted with a combined global local reinforcement method, then upon goal achievement, the hidden layer will contain information that can be used to plan for that goal. This hypothesis has begun to be tested in subsequent work [25]. The hidden layer representations at the time of goal achievement were termed protoplans. To investigate whether protoplans could actually help to guide behavior, a transfer of learning experiment was done. The protoplans learned in one controller network were used to guide a second network as it ....

L. A. Meeden, Towards planning: Incremental investigations into adaptive robot control. PhD dissertation, Indiana University, 1994.

No context found.

Meeden, L. (1994). Towards Planning: Incremental Investigations into Adaptive Robot Control. Unpublished Ph.D. dissertation, Indiana University, Department of Computer Science.

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