Henrik Hautop Lund and John Hallam. Sufficient neurocontrollers can be surprisingly simple. Research paper no. 826. Department of Artificial Intelligence, University of Edinburgh, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....been implemented, it can potentially automaticly produce controllers for related tasks throught only very minor changes. The main advantages of the approach is the potential to reduce number of man hours used to construct the system, and the potential of the system to produce innovative solutions [8] to a given problem. Resulting Context: As a result of applying the solution of this pattern, you have a system that in theory should generate robot controllers for the task at hand. You might however find, that it is infeasible to use the system due to the time required, wear and tear of the ....

Henrik Hautop Lund and John Hallam. Sufficient neurocontrollers can be surprisingly simple. Research paper no. 826. Department of Artificial Intelligence, University of Edinburgh, 1996.

.... same (i.e. number of individuals, mutation rate, lifetime length, architecture of control system, genotype tophenotype mapping, coding of the sensory motor information, robot body) quite different forms of behavior can be obtained: a) obstacle avoidance (Floreano Mondada, 1994) b) exploration (Lund Hallam, 1996); c) navigation (Nolfi 14 Miglino, in press) d) discrimination of different objects (Nolfi, 1997c) e) escaping predators (Nolfi Floreano, in press) Of course, some of the characteristics which could be included in the evolutionary process but are often set by the experimenter (in ....

Lund, H. H. & Hallam, J. (1996) Sufficient neurocontrollers can be surprisingly simple.

....which allows us, using a limited amount of supervision, to leave the system free to develop its own strategy to solve the task. We then analyze the solutions found by the evolutionary process. Finally, we discuss the implication of the results obtained. 2. Exploring an arena surrounded by walls Lund Hallam (1996) reported a simulation in which a Khepera robot was trained to perform an exploratory task (i.e. to visit as much of the environment as possible within a limited time) We present the results of a replication of this experiment in almost identical conditions. 5 Khepera is a miniature mobile ....

.... As in Lund and Hallam s experiments, this is accomplished by using a specific turning angle each time the robot approaches a wall, so that the perceptron 6 navigates the robot through the environment exactly so as to have the robot encountering previously non visited cells most of the times (Lund Hallam, 1996: 11) Figure 1. Simulated trajectory of a typical evolved individual left free to move in the arena for 1500 cycles (150 seconds) Almost identical behavior can be observed by downloading the individual s control system into the real robot and testing it in the real environment (see Lund ....

[Article contains additional citation context not shown here]

Lund, H. H., & Hallam, J. H. 1996. Sufficient neurocontrollers can be surprisingly simple. Technical Report, University of Edinburgh, Department of Artificial Intelligence.

....perceptrons, for example, use the same sigmoid function of equation 2.2 but have only two layers of units: one for inputs and one for outputs. Not only have these been used by Floreano and Mondada (1994) to perform the same obstacle avoidance behaviour on a Khepera robot as mentioned above, but Lund and Hallam (1996) have shown that they are capable of evolving behaviours such as exploration and homing that, at first sight, one might not even think of as reactive. Other types of networks that have been used to evolve reactive behaviours include biologically inspired continuous time networks (Hopfield 1984) as ....

Lund, H. and J. Hallam (1996). Sufficient neurocontrollers can be surprisingly simple. Research Paper 824, Department of Artificial Intelligence, University of Edinburgh.

....by biologists. Other experiments, where a description of the robot trajectory is crucial, include our Evolutionary Robotics experiments where we evolve robot controllers (e.g. neurocontrollers) and or robot body plans in a simulator before transferring the evolved systems to real robots, e.g. [3, 2]. In these Evolutionary Robotics experiments it is crucial to show the nature of the mapping between the behavior in simulation and reality. The system used for the purpose of recording the track of mobile robots Technical Paper no. 41, Department of Artificial Intelligence, University of ....

H. H. Lund and J. Hallam. Sufficient Neurocontrollers can be Surprisingly Simple. Research Paper 824, Department of Artificial Intelligence, University of Edinburgh, 1996.

....locomotion and steering and one passive trolley wheel. The robot received sensory information from 4 mechanical switches mounted within two bumpers, 2 ambient light sensors at its front, and 2 reflection detectors under the chassis. Right: Photo of the LEGO robot. c fl Lund et al. 1997. ported in [14, 16, 22], it is possible to build an accurate simulator that allows one to go from simulation to reality by basing the simulator on the robot s own samplings of sensor and motor responses. The LEGO simulator was built by this technique (described in section 2) The architecture of the control system that ....

H. H. Lund and J. Hallam. Sufficient Neurocontrollers can be Surprisingly Simple. Research Paper 824, Department of Artificial Intelligence, University of Edinburgh, 1996.

No context found.

Lund, H.H. and Hallam, J. Sufficient Neurocontrollers can be Surprisingly Simple. Research Paper 824. Department of Artificial Intelligence. Edinburgh University. 1996.

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