FENS Forum 2010 - Amsterdam - Posters To be on display from 8:00 to 13:15 in the morning and from 13:30 to 18:45 in the afternoon. Poster sessions run from 09:30 to 13:15 in the morning and from 13:30 to 17:30 in the afternoon. A one hour time block is dedicated to discussion with the authors (authors should be in attendance at their posters as from the time indicated.) - For other sessions, time indicates the beginning and end of the sessions.

Poster board F30 - Tue 06/07/2010, 13:30 - Hall 1
Session 176 - Learning 4
Abstract n° 176.30
Publication ref.: FENS Abstr., vol.5, 176.30, 2010

Authors	Lakhman K. V. (1) & Burtsev M. S. (2)
Addresses	(1) Nat. Res. Nucl. Univ. "MEPhI", Moscow, Russia; (2) Non-linear Dynamics Inst. Appl. Math. RAS, Moscow, Russia
Title	The neural network model of evolution and learning of goal-directed behaviour
Text	Animal behaviour is determined by the need to achieve different goals. Often goals can be decomposed into a number of interim tasks. Such goal-directed behaviour consists of innate and learned components. It identifies the importance of computer simulation of evolution and learning of goal-directed behaviour. We simulated an evolution in the population of learning autonomous agents. The environment, where the population of agents evolved, was represented as a binary vector. Every time the agent took a step it modified a single bit of environment. The environment contained a number of competitive goals and subgoals of varying complexity, that were defined as an ordered set of changes of the state vector. Each goal was associated with certain reward value directly proportional to its complexity. The reproductive success of the agent was proportional to the accumulated reward. The agent's behaviour was controlled by a multilayer neural network with arbitrary topology (including recurrent connections) modified by evolution and then learning. During lifetime of the agent each neuron in neural network formed a prediction of expected afferentation. Learning arises through the formation of new functional neuronal groups by gradual inclusion of "silent" neurons if the goal cannot be fulfilled. The simulations for the stationary and quasistationary environments demonstrated that in the latter a much wider range of possible behavioural policies and higher cumulative rewards were observed. We conclude that in variable environments agents with the behavior, allowing them to reach a greater number of goals, have an evolutionary advantage. However, with increasing instability of the environment over a certain level we observed a sharp decline in the efficiency of evolutionary adaptation. By increasing a number of goals and the growing complexity of its structure higher quality of learning can be achieved. Moreover, the analysis of internal dynamics of the evolutionary adaptation was done.
Theme	F - Cognition and behaviour Animal cognition and behaviour / Cognitive learning and memory systems

Close window