Cognitive Control: Theory and Application

Mehdi Fatemi

doi:10.1109/ACCESS.2014.2332333

2014-01-01

Mehdi Fatemi *

S. Haykin

DOI：10.1109/ACCESS.2014.2332333

原文链接

查看原文

摘要

En 中文

From an engineering point-of-view, cognitive control is inspired by the prefrontal cortex of the human brain; cognitive control may therefore be viewed as the overarching function of a cognitive dynamic system. In this paper, we describe a new way of thinking about cognitive control that embodies two basic components: learning and planning, both of which are based on two notions: 1) two-state model of the environment and the perceptor and 2) perception-action cycle, which is a distinctive characteristic of the cognitive dynamic system. Most importantly, it is shown that the cognitive control learning algorithm is a special form of Bellman's dynamic programming Distinctive properties of the new algorithm include the following: 1) optimality of performance; 2) algorithmic convergence to optimal policy; and 3) linear law of complexity measured in terms of the number of actions taken by the cognitive controller on the environment. To validate these intrinsic properties of the algorithm, a computational experiment is presented, which involves a cognitive tracking radar that is known to closely mimic the visual brain. The experiment illustrates two different scenarios: 1) the impact of planning on learning curves of the new cognitive controller and 2) comparison of the learning curves of three different controllers, based on dynamic optimization, traditional Q-learning, and the new algorithm. The latter two algorithms are based on the two-state model, and they both involve the use of planning.

Keyword:

Cognitive dyanamic systems

cognitive control

dynamic programming

two-state model

entropic state

Shannon's entropy

explore/exploit tradeoff

learning

planning

Bayesian filtering