Toggle Main Menu Toggle Search

Open Access padlockePrints

Natural dynamics and neural networks: Searching for efficient preying dynamics in a virtual world

Lookup NU author(s): Dr Peter Andras


Full text for this publication is not currently held within this repository. Alternative links are provided below where available.


Since the 1970s, the chaotic nature of environmental dynamics of animals is a challenge for computer-science researchers. The intriguing question is how can we develop an effective means for predicting future states? In this paper we assume that animals possess control structures exhibiting chaotic dynamics that can deal with chaotic environmental dynamics. Based on that assumption, we describe a novel model that is characterized by predicting future states through a recursive application of an input-output mapping generated by natural dynamics (the internal chaotic dynamics) and chaotic environmental dynamics. For comparison, we evaluate two neural networks that control the prediction behavior of an agent (for example, an animal) in a dynamical environment. The inspirational source of our simulated environment is the preying dynamics of the frog-fly interactions. First we examine a non-chaotic direct-prediction network that predicts future states by mapping the input in a single step to the output. Then we develop and investigate a chaotic recursive-prediction network that predicts the future in a recursive manner, using iterations of small-step predictions. Our experiments show that the recursive-prediction network clearly outperforms the direct-prediction network in the sense of the number of accurate predictions. The experimental tasks involved the prediction of future states of a chaotically moving target. The result is also explained in terms of complexity of the underlying function-approximation task. From our observations that recursive-prediction networks are better than direct-prediction networks in dealing with chaotic environments, and from the performance of the recursive-prediction network, we tentatively conclude that our model can serve as a chaotic-deterministic component in the procedure of predicting future states.

Publication metadata

Author(s): Andras P, Postma E, Van Den Herik J

Publication type: Article

Publication status: Published

Journal: Journal of Intelligent Systems

Year: 2001

Volume: 11

Issue: 3

Pages: 173-202

Print publication date: 01/01/2001

ISSN (print): 0334-1860

ISSN (electronic):

Publisher: Freund Publishing House