Adaptive capture of expert behavior
The authors smoothed and captured a set of expert rules with adaptive networks. The motivation for doing this is discussed. (1) Smoothing leads to stabler control actions. (2) For some sets of rules, the evaluation of the rules can be sped up. This is important in large-scale simulations where many intelligent elements are present. (3) Variability of the intelligent elements can be achieved by adjusting the weights in an adaptive network. (4) After capture has occurred, the weights can be adjusted based on performance criteria. The authors thus have the capability of learning a new set of rules that lead to better performance. The set of rules the authors chose to capture were based on a set of threat determining rules for tank commanders. The approach in this paper: (1) They smoothed the rules. The rule set was converted into a simple set of arithmetic statements. Continuous, non-binary inputs, are now permitted. (2) An operational measure of capturability was developed. (3) They chose four candidate networks for the rule set capture: (a) multi-linear network, (b) adaptive partial least squares, (c) connectionist normalized local spline (CNLS) network, and (d) CNLS net with a PLS preprocessor. These networks were able to capture the rule set to within a few percent. For the simple tank rule set, the multi-linear network performed the best. When the rules were modified to include more nonlinear behavior, CNLS net performed better than the other three nets which made linear assumptions. (4) The networks were tested for robustness to input noise. Noise levels of plus or minus 10% had no real effect on the network performance. Noise levels in the plus or minus 30% range degraded performance by a factor of two. Some performance enhancement occurred when the networks were trained with noisy data. (5) The scaling of the evaluation time was calculated. (6) Human variation can be mimicked in all the networks by perturbing the weights.
- Research Organization:
- Los Alamos National Lab., NM (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- W-7405-ENG-36
- OSTI ID:
- 10179483
- Report Number(s):
- LA-UR--94-2825; ON: DE94018285
- Country of Publication:
- United States
- Language:
- English
Similar Records
Adaptive capture of expert knowledge
Adaptive information retrieval: machine learning in associate networks
Using CNLS-net (Connectionist Normalized Local Spline-network) to predict the Mackey-Glass chaotic time series
Technical Report
·
Mon May 01 00:00:00 EDT 1995
·
OSTI ID:93561
Adaptive information retrieval: machine learning in associate networks
Thesis/Dissertation
·
Tue Dec 31 23:00:00 EST 1985
·
OSTI ID:5271568
Using CNLS-net (Connectionist Normalized Local Spline-network) to predict the Mackey-Glass chaotic time series
Conference
·
Mon Dec 31 23:00:00 EST 1990
·
OSTI ID:5749339