Baofu Duan

Iterative feature weighting for identification of relevant features with radial basis function networks

This paper reports on advances in identification of relevant features through iterative feature weighting with radial basis function networks. It proceeds with a set of feature weights to scale the data which are used to train a radial basis function network model. Then from the learned model, the feature weights are updated via one-step gradient descent. The updated feature weights are then fed back to build a new model. The procedure continues until we find a satisfactory model and the feature weights converge. Experimental results for some benchmark datasets show that the approach is efficient and effective for selecting relevant features for data modeling and classification tasks.

Implementing Variable Strategy in Real-Time Reflex Decision-Making and Control

Abstract In principle, autonomous decision-making and control can be implemented efficiently in reflex mode with use of neural networks. In practice there are situations where the very strategy of decision or control needs to be varied in accordance with circumstances. For example, in multi-objective control, emphasis may need to be shifted from one objective to another in accordance with system state. Similar considerations apply to decision-making. This paper reports on how knowledge of such variation in strategy can be captured and implemented. Control of the inverted-pendulum/cart task is used to illustrate the approach. Other applications are also discussed.

Neural-Network Based Exploration in the Discovery and Design of Materials

Abstract In previous work we reported on materials design and property estimation using empirical or ordinal representation approaches (Pao et al., 1999, Pao et al., 2000). In this paper, we report in detail on a Basic Concepts approach, which combines some of the merits of both the ab initio quantum theoretic first principles approach and the commonly used empirical approach. We show that it is possible to establish mappings between sets of basic atomic characteristics and properties of materials systems formed from those constituent atomic species. Such mappings are implemented in neural networks and are trained using large bodies of high-quality well-filtered data. We also report on the adaptive extrapolation and exploration of such neural network mappings. Such exploration can be used to suggest likely compositions of new compounds that would have certain required properties. This is accomplished by a combined technique of neural networks and evolutionary search. The feasibility and the effectiveness of this method are illustrated with examples.

Evolving a Juggler: Using the Inverted Pendulum/Cart Control Task to Investigate the Issue of How to Modulate Reflexes

Abstract Paper reports on an investigation of how adaptive intelligent reflex controllers can be trained through an evolutionary process using recurrent coupled neural nets acting in direct control manner. The motivation is to explore how reflex actions are learned and how reflex action might be modulated adaptively. Two tasks are discussed. Results indicate that ‘intelligent’ reflex behavior can be implemented through coupling of dynamical systems. This mode of control seems to be eminently suitable for exercising adaptive real-time control.