Evolving Dynamical Neural Networks for Adaptive Behavior

Abstract

We would like the behavior of the artificial agents that we construct to be as well-adapted to their environments as natural animals are to theirs. Unfortunately, designing controllers with these properties is a very difficult task. In this article, we demonstrate that continuous-time recurrent neural networks are a viable mechanism for adaptive agent control and that the genetic algorithm can be used to evolve effective neural controllers. A significant advantage of this approach is that one need specify only a measure of an agent's overall performance rather than the precise motor output trajectories by which it is achieved. By manipulating the performance evaluation, one can place selective pressure on the development of controllers with desired properties. Several novel controllers have been evolved, including a chemotaxis controller that switches between different strategies depending on environmental conditions, and a locomotion controller that takes advantage of sensory feedback if available but that can operate in its absence if necessary.

Neuroevolution of Feedback Control for Object Manipulation by 3D Agents

“…We use a bespoke, hybrid neural network (HNN) to this end. The HNN comprises feed-forward networks for saliency calculation from sensor data, a locallyconnected, topologically-organised shunting neural network (Yang and Meng, 2000) for modelling the agent's world, a feed-forward bridge between this model and the motor control parts of the architecture and a series of recurrent leaky integrator networks in the style of Beer's Continuous-Time RNNs (Beer and Gallagher, 1992) that actually produce motor output from the control system. These components we label the decision network (DN), the shunting model (SM), the physical network (PN) and the pattern generators (PG).…”

Embodied artificial life at an impasse can evolutionary robotics methods be scaled?

“…Early on, Harvey et al studied locomotion with object avoidance [63] (also see [48]). Zone homing behaviors were evolved in [42]. Locomotion with homing was investigated in [46,47].…”

“…One work in particular, performed by Friedman [37] in the 1950's, used a form of artificial evolution to evolve simulated robots to perform a chemo-gradient following task. This work foreshadowed by more than thirty years the rise of situated agent-based artificial life [38][39][40][41] and its embodied counterpart, evolutionary robotics [42][43][44][45][46][47][48][49].…”

“…The robot controllers must evolve all low level competencies, including object avoidance and ball collection and delivery, hence the task includes a significant level of behavior acquisition and integration. The most complex ER tasks may fall into the range of minimal cognition as set out by Beer [3]. However, all ER control tasks studied to date can be solved by relatively simple algorithms.…”

Evolution of Signalling in a Group of Robots Controlled by Dynamic Neural Networks

“…Given that the task we want to study requires the use of time-dependent structures, we use fully connected, thirteen neuron Continuous Time Recurrent Neural Networks (CTRNN's see [10])-see Fig. 3 for a depiction of the network.…”