F. Pasemann. Evolving neurocontrollers for balancing an inverted pendulum. Network: Comput. Neural Syst., 9:495 511, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....potential corresponding to at followed by a slower relaxation with respect to a2. The two muscle tensions are summed to provide a force F to move the cart which balances a pendulum on it. Therefore the muscles together with the cart pendulum arrangement described by the standard model found in [8, 9] is the highly non linear system to be controlled. We provide as feedback the position of the pendulum as 0, the velocity p and the force F. The goal is to move the cart, such that the pendulum position follows the reference input I(t) 0r,f (t) This task is, due to the intermediate muscles, ....

....behaviour prescribed by (13) and also in this case it can be arbitrary fast. Fig. 3(a) c) show the performance of the complete model. In (a) the pendulum is regulated to rest at 0 = 0 which is obtained very fast in m 0. 15s, which is extremely efficient compared to the results of for instance [9], where typical times are at leat m 2s (for easier tasks) In (b) the fast control actions are shown, which drive the pendulum to follow a reference 0.15 sin(10t) in the presence of a faster sinusoid disturbance. In (c) we added more erratic fast noise. In all cases the system is perfectly ....

F. Pasemann. Evolving neurocontrollers for balancing an inverted pendulum. Network: Comput. Neural Syst., 9:495 511, 1998.

.... algorithms turned out to be a very e ective tool for general problem solving [33] It was also shown to generate interesting classes of robot behaviors (see e.g. 23] Instead of standard genetic algorithms, here an ENS 3 algorithm (evolution of neural systems by stochastic synthesis) [27] is used. It is applied to networks of standard additive neurons with sigmoidal transfer functions and sets no constraints neither on the number of neurons nor on the connectivity structure of networks. It develops network architecture and optimizes parameters like weights and bias terms ....

....and bias terms simultaneously. In acquiring both network topology and parameter optimization at the same time it is similar to the GNARL algorithm [2] In contrast to genetic algorithms it does not quantize network parameters, and it was tested successfully for some non linear control problems [27]. For the solution of extended problems (more complex environments, more complex sensori motor systems, more complex survival conditions, etc. the synthesis of evolved neuromodules forming larger neural systems can be achieved by evolving the coupling structure between modules. Of course, ....

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495{ 511.

....are also tested on a physical environment with the use of Khepera robots. 1 1 Introduction Within the frame of Synthetic Modeling and Embodied Cognitive Science [4] an evolutionary algorithm is presented. This algorithm, called ENS 3 for Evolution of Neural Systems by Stochastic Synthesis [7]) evolves the structure and size of artificial neural networks and optimizes the corresponding parameters at the same time. It is designed especially to generate networks with recurrent connectivity. The starting hypothesis for the presented experiments is that internal dynamical properties of ....

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495--511. -- 9

....with high performance are also tested on a physical environment with the use of Khepera robots. 1 1 Introduction Within the frame of Synthetic Modeling and Embodied Cognitive Science [8] an evolutionary algorithm is presented. The Evolution of Neural Systems by Stochastic Synthesis (ENS 3 ) [7], evolves the structure and size of artificial neural networks and optimizes the corresponding parameters at the same time. It is designed especially to generate networks with recurrent connectivity. In acquiring both network topology and parameter optimization simultaneously it is similar to the ....

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495--511.

....[1] an evolutionary algorithm to develop neuromodules as well as the couplings between these subsystems is used. This algorithm is called ENS 3 algorithm for evolution of neural systems by stochastic synthesis. The ENS 3 has already been tested successfully on nonlinear control problems [6]. Furthermore, autonomous systems acting in a sensori motor loop, like simulated or real robots, are an appropriate tool for studying the development of embodied cognition [7] For the experiments reported in this paper the miniature Khepera robots [3] as well as the Khepera simulator [2] are ....

.... coded as follows F : 2000 X t=3D1 ff Delta in f (t) fi Delta j in f (t) Gamma in r (t) j ; 2) where ff and fi are appropriate parameters, and in f and in r are given in terms of the inputs i[0] i[7] to the additional light sensors by in f : i[0] i[2] i[3] i[5] in r : i[6] i[7] Thus, for determining the fitness only four of the six front light sensors are used, and the proximity sensors are not evaluated at all. Again, the ENS 3 algorithm is applied to the simulated robots with a few light sources now spread over the environments used also in the first ....

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495--511.

....forming larger neural systems can be achieved by evolving the coupling structure between modules. This is done in the spirit of coevolution of interacting species. We suggest that this kind of evolutionary computation is better suited for evolving neural networks than genetic algorithms. In [7] we reported on tests of the algorithm, applying it to the pole balancing problem that usually serves as a benchmark problem for trainable controllers [5] Of course, the inverted pendulum is one of the simplest inherently unstable systems, and balancing it under benchmark conditions is mainly in ....

....failure before the end of the maximal evaluation time t max = 12 seconds. The module in figure 1 displays already an interesting feature: it can be understood as composed of two submodules. The structure of the one given by neuron 6 with its four inputs is known as that of a pole balancing module [7]. The module given by neuron 5 with its three inputs swings up the pole from downward positions when isolated. They are coupled through the connection w 56 . 3.2 An s class Controller with Four Inputs Sensor signals are again given by equation (3) The force on the cart is applied according to ....

[Article contains additional citation context not shown here]

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495--511.

....forming larger neural systems can be achieved by evolving the coupling structure between modules. This is done in the spirit of coevolution of interacting species. We suggest that this kind of evolutionary computation is better suited for evolving neural networks than genetic algorithms. In [7] we reported on tests of the algorithm, applying it to the pole balancing problem that usually serves as a benchmark problem for trainable controllers [5] Of course, the inverted pendulum is one of the simplest inherently unstable systems, and balancing it under benchmark conditions is mainly in ....

....failure before the end of the maximal evaluation time t max = 12 seconds. The module in gure 1 displays already an interesting feature: it can be understood as composed of two submodules. The structure of the one given by neuron 6 with its four inputs is known as that of a pole balancing module [7]. The module given by neuron 5 with its three inputs swings up the pole from downward positions when isolated. They are coupled through the connection w 56 . 3.2 An s class Controller with Four Inputs Sensor signals are again given by equation (3) The force on the cart is applied according to ....

[Article contains additional citation context not shown here]

Pasemann, F. (1998), Evolving neurocontrollers for balancing an inverted pendulum, Network: Computation in Neural Systems, 9, 495-511.

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