1 |
Learning in fractured problems with constructive neural network algorithmsKohl, Nate F. 23 March 2011 (has links)
Evolution of neural networks, or neuroevolution, has been a successful approach to many low-level control problems such as pole balancing, vehicle control, and collision warning. However, certain types of problems — such as those involving strategic decision-making — have remained difficult to solve. This dissertation proposes the hypothesis that such problems are difficult because they are fractured: The correct action varies discontinuously as the agent moves from state to state. To evaluate this hypothesis, a method for measuring fracture using the concept of function variation of optimal policies is proposed. This metric is used to evaluate a popular neuroevolution algorithm, NEAT, empirically on a set of fractured problems. The results show that (1) NEAT does not usually perform well on such problems, and (2) the reason is that NEAT does not usually generate local decision regions, which would be useful in constructing a fractured decision boundary. To address this issue, two neuroevolution algorithms that model local decision regions are proposed: RBF-NEAT, which biases structural search by adding basis-function nodes, and Cascade-NEAT, which constrains structural search by constructing cascaded topologies. These algorithms are compared to NEAT on a set of fractured problems, demonstrating that this approach can improve performance significantly. A meta-level algorithm, SNAP-NEAT, is then developed to combine the strengths of NEAT, RBF-NEAT, and Cascade-NEAT. An evaluation in a set of benchmark problems shows that it is possible to achieve good performance even when it is not known a priori whether a problem is fractured or not. A final empirical comparison of these methods demonstrates that they can scale up to real-world tasks like keepaway and half-field soccer. These results shed new light on why constructive neuroevolution algorithms have difficulty in certain domains and illustrate how bias and constraint can be used to improve performance. Thus, this dissertation shows how neuroevolution can be scaled up from learning low-level control to learning strategic decision-making problems. / text
|
Page generated in 0.012 seconds