Paper Info
Title
Analysis of a Dynamical Recurrent Neural Network Evolved for Two Qualitatively Different Tasks: Walking and Chemotaxis
Abstract
Living organisms perform a broad range of different be- haviours during their lifetime. It is important that these be coordinated such as to perform the appropriate one at the right time. This paper extends previous work on evolving dynami- cal recurrent neural networks by synthesizing a single circuit that performs two qualitatively different behaviours: orien- tation to sensory stimuli and legged locomotion. We demon- strate that small fully interconnected networks can solve these two tasks without providing a priori structural modules, ex- plicit neural learning mechanisms, or an external signal for when to switch between them. Dynamical systems analy- sis of the best-adapted circuit explains the agent's ability to switch between the two behaviours from the interactions of the circuit's neural dynamics, its body and environment. isolation. Such a divide-and-conquer approach can be very useful for engineering robots that need to perform multiple complex tasks, not least because it simplifies the understand- ing of how the robot works. But it is less useful in the context of developing the tools and language to understand biologi- cal organisms, as these may not necessarily have evolved to be easily decomposable. First, we investigate whether a single neurocontroller can exhibit qualitatively different behaviours without imposing constraints on its structure. We use artificial evolution to synthesize a recurrent neural network that when coupled to two different simulated bodies, namely a one legged insect and a two-wheeled robot with a chemical sensor, has to per- form legged locomotion in the former and chemotaxis in the latter case1. A successful agent has to detect which body it inhabits and generate the appropriate behaviour. It must do this in the absence of an external signal and without any on- line changes in the parameters of the controller. We aim to find the smallest network that can solve the task. Although the structure of the network is under evolution, we do not investigate whether the evolved networks exhibit a degree of structural or 'functional'2 modularity. Second, using the mathematical tools of dynamical sys- tems theory, we explain how the circuit in interaction with its body and environment can generate distinct behaviours. We characterize the autonomous dynamics of the best-evolved circuit and how its dynamics vary with inputs. We then study how the observed behavioural patterns are generated through the closed-loop interaction of the neural dynam- ics with the body and environment, for the two different tasks. Finally, we show how the evolved agent makes use of context-dependent feedback to shape the different transients using the same dynamical landscape. This leads us to sug- gest a dynamical systems perspective on adaptive behaviour that goes beyond attractors.
Year
Venue
Keywords
2008
ALIFE
dynamic system,artificial evolution,context dependent,recurrent neural network,divide and conquer
Field
DocType
Citations 
Attractor,Control theory,Evolutionary algorithm,Computer science,A priori and a posteriori,Recurrent neural network,Dynamical systems theory,Artificial intelligence,Robot,Machine learning,Modularity
Conference
9
PageRank 
References 
Authors
0.63
9
3
Name
Order
Citations
PageRank
Eduardo Izquierdo1467.91
Thomas Buhrmann290.97
Birmingham B3151.24