S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured rnvironment. Connection Science 11, 2:129--152, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....and another neural network(NN) to select the correct neural network to use. EM is used to associate each data point (sensor reading) with one of the NN experts. The maximization step corresponds to normal neural network training. This technique is trained offline and predictive. Nolfi and Tani [32] propose a technique of cascaded NNs. The lowest level NN tries to predict the next sensor reading. The hidden units of this NN are then clustered into states by a neural network. The clustered hidden states are then used to train a higher level NN which attempts to learn the state transitions of ....

..... # . # . 12 Table 3: Neural Network Techniques Description DOLCE Gated Experts Cascaded NN Neural Navigation Reference [5] 38] [32] [34] First Author Das Weigend Nolfi Tani . Parameterless Data properties . Probability distributions . Information missing ....

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured rnvironment. Connection Science 11, 2:129--152, 1999.

.... presented here are based on actor critic RL [8] The planning components, the planning algorithm, and coarse planning are new, but the idea of planning as a RL system trained within a model of the environment is from [5] cf. 4] for a NN implementation of the Dyna architectures and [4] [6] for a NN implementation of a model of the environment) Section 2 illustrates the simulations scenario and the NN planning controller. Section 3 presents the results and their discussion. Section 4 draws some conclusions. 2 Scenario of Simulations and Controller The environment used in the ....

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured environment. Connection Science, 2(11):129-152, 1999.

....(Figure 2) is only updated when an event extraction mechanism signals that a change in inputs has been detected. In between these events, this higher level (echelon 2 system) remains in a suspended state. Thereby the system becomes decoupled from the immediate here andnow of the robot s operation [4]. Realization Sensation Unsupervised Classifier Event Action (asynchronous) synchronous) Behavior Figure 2: Overview of the layered architecture which can deal with arbitrarily long delay periods between relevant stimuli. Echelon 1 performs real time control of the system; its ....

S. Nolfi and J. Tani, "Extracting regularities in space and time through a cascade of prediction networks," Connection Science, Vol. 11:2, pp. 125--148, 1999

.... is, the ARAVQ Learning Delayed Response Tasks through Unsupervised Event Extraction 9 0 7 2 1 left motor right motor 4 3 6 else motor[LEFT] 5; motor[RIGHT] 5; motor[LEFT] 5; else if (sensor[1] 800) motor[RIGHT] 0; motor[LEFT] 0; motor[RIGHT] 5; if (sensor[4] 800) Corridor follower motor[LEFT] 5; motor[RIGHT] 4; motor[LEFT] 2; motor[RIGHT] 5; else motor[LEFT] 5; motor[RIGHT] 5; else if (sensor[0] 800) Left wall follower else motor[LEFT] 5; motor[RIGHT] 5; if (sensor[4] 200) ....

.... = 5; if (sensor[4] 800) Corridor follower motor[LEFT] 5; motor[RIGHT] 4; motor[LEFT] 2; motor[RIGHT] 5; else motor[LEFT] 5; motor[RIGHT] 5; else if (sensor[0] 800) Left wall follower else motor[LEFT] 5; motor[RIGHT] 5; if (sensor[4] 200) motor[LEFT] 4; motor[RIGHT] 5; else if (sensor[5] 800) motor[LEFT] 5; motor[RIGHT] 2; Right wall follower (c) d) Fig. 5. The Khepera robot and the hand crafted behaviors. network received a total of 12 inputs. The parameter settings of the ARAVQ were = ....

S. Nol and J. Tani. Extracting regularities in space and time through a cascade of prediction networks. Connection Science, 11(2):125-148, 1999.

.... paper presents is simple: the clusters found at a hidden layer of a recurrent network can be segmented and discretized using a self organizing map (or other unsupervised classifier) and be used as inputs and targets of another recurrent network, similarly trained to predict the next input (cf. [7] who used a cascaded setup to abstract from low level events in a robotic environment, and [10] who trained, on a variety of domains, an extra network to produce class representations (out of target data) which the main network is able to predict) Effectively, through continuous training, the ....

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured environment. Connection Sci, 11(2):125--148, 1999.

....account in form of an asynchronous event stream, while the operation of the robot sensors and actuators work on a synchronous time step based stream of sensations and actions. The problem of getting from time stepped synchronous input streams to asynchronous event streams, has been addressed in [8,5] where layered bottom up approaches were employed. Both these systems, however, had in common that they did not involve any way of getting back from an event stream to the level of time step based inputs and outputs, i.e. the robots could not use the events to give a response; a critical part in ....

....could be incorporated at any time point. Every other part of the system that is to receive these outputs would need to have likewise growing frontend. Despite the orthogonality and the growing back end problems, such binary localistic representations have been used for encoding events e.g. in both [8] and [4] A solution to the orthogonality problem would be to skip the winner take all aspect involved in classification thereby allowing real valued activation, but still under the localistic regime. That is, each class unit is activated to the degree it matches the input [3] This has much in ....

[Article contains additional citation context not shown here]

Nolfi, S. and Tani, J. Extracting regularities in space and time through a cascade of prediction networks. Connection Science, 11(2):125--148, 1999.

.... The idea this paper presents is simple: the clusters found at a hidden layer can be segmented and discretized using a self organizing map (or other unsupervised classifier) and be used as inputs and targets of another recurrent network, similarly trained to predict the next input (cf. [7] who used a cascaded setup to abstract from low level events in a robotic environment, and [10] who trained, on a variety of domains, an extra network to produce class representations (out of target data) which the main network is able to predict) Effectively, through continuous training, the ....

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured environment. Connection Sci, 11(2):125--148, 1999.

....Our solution is based on having several information processing levels, as depicted in Figure 2. Level 1: Contains raw multi dimensional sensor data, which is time step based 1 . This is the level where [Rylatt and Czarnecki, 2000] approached the delayed response task. We however argue, like [Nolfi and Tani, 1999] , that real world 1 Where a time step is defined as a single update of sensors, neuronal elements, and actuators with a regular time interval, typically in the millisecond range. a a a a b a c d e Classification aaabbbbbaaaaaaaacccccaaaaddddaaaaaaaaaaaaaaaaaaaaaeeeeeaaaaaa Event filtering b ....

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks. Connection Science, 11(2):125--148, 1999.

....sensory and motor systems. We also show how this technique can be used to construct global maps from an abstract description of the sensory ow. 1. Introduction We present a technique for analysis of abstract sensory ows based on sensor and motor inversion. As shown by (Tani and Nol , 1998, Nol and Tani, 1999), sensory ows of mobile robots can be automatically segmented and represented in a compact manner as a series of extracted (more abstract) higher level concepts such as wall, right turn, corridor, left turn, etc. Figure 1) 6 Figure 1: The sensory ow of a mobile robot can be split into ....

....they are manifested. A simple solution is to use an absolute technique, which divides the input space into regions using a set of absolute borders. No matter how slowly such a border is crossed, it can be detected. Nol and Tani used two di erent architectures, described in (Tani and Nol , 1998, Nol and Tani, 1999), which in e ect did this. Further, Nol and Tani showed that the sensory ow was automatically split into segments which, to a distal observer, actually corresponded to concepts such as corridor, wall, corner, etc. 1.2 Segmentation principles Both of Nol and Tani s architectures were based on ....

[Article contains additional citation context not shown here]

Nol, S. and Tani, J. (1999). Extracting Regularities in Space and Time Through a Cascade of Prediction Networks: The Case of a Mobile Robot Navigating in a Structured Environment. Connection Science 11(2): 125-148.

....between them. Labelling each segment with a symbol, the entire sequence can be stored (with some loss of information) using just seven symbols instead of hundreds, or even thousands, of distinct input values. The length of each sub sequence is however lost during the mapping Nol and Tani [7, 9] conducted experiments using a similar wall following robot and segmented the sensory ow using a hierarchical neural network architecture consisting of several prediction and segmentation networks. While their system managed to extract higher order concepts from the sensory ow, such as walls , ....

....vector quantizer, which solves these problems. It is de ned mathematically and results of the experiments conducted with this architecture are presented. Finally, in section 4, the main advantages of the new approach are summarized. 2 Existing Methods In experiments carried out by Nol and Tani [7, 9], di erent neural network architectures were investigated which segmented sensory input sequences from mobile robots. In the former paper [9] they designed a modular system of gated experts, where each module represented a sub sequence. In the latter paper [7] they presented an altered ....

[Article contains additional citation context not shown here]

Nol S., Tani J.: Extracting Regularities in Space and Time Through a Cascade of Prediction Networks: The Case of a Mobile Robot Navigating in a Structured Environment, Connection Science (1999), 11(2).

No context found.

Nolfi, S. & Tani, J. (1999). Extracting regularities in space and time through a cascade of prediction networks: the case of a mobile robot navigation in a structured environment. to appear.

No context found.

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured rnvironment. Connection Science 11, 2:129--152, 1999.

No context found.

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured environment. Connection Science, 11(2):1131--1141, 1999.

No context found.

Stefano Nolfi and Jun Tani. Extracting regularities in space and time through a cascade of prediction networks: The case of a mobile robot navigating in a structured environment. Connection Science, 11(2):125--148, 1999.

No context found.

S. Nolfi and J. Tani. Extracting regularities in space and time through a cascade of prediction networks: the case of a mobile robot navigation in a structured environment, 1999.

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