|
Computational modeling of learning in complex problem solving tasks| old_uid | 4395 |
|---|
| title | Computational modeling of learning in complex problem solving tasks |
|---|
| start_date | 2008/03/21 |
|---|
| schedule | 11h |
|---|
| online | no |
|---|
| summary | Information processing theory emphasizes search and heuristics in problem solving and the role of learning has been relatively neglected. In my doctoral work, I address how humans learn to solve problems under three realistic learning conditions and present computational simulations that model that learning. Experimental results showed that people solved the problem more accurately after
learning from demonstrations (imitation learning) or verbal instructions than from binary feedback indicating only whether answers were correct or not (reinforcement learning). The binary feedback may not be necessary because people were able to evaluate solutions on their own. When learning from demonstrations, people generalized to more complex tasks, suggesting understanding rather than rote memorizing of the observed solutions. Computational models of reinforcement-learning were less accurate than humans, supporting the claim that humans use self-evaluation to learn the task. Models of imitation learning and verbal instructions were as accurate as humans.
I also present ideas for a new, parsimonious model that (1) learns from both demonstrations and from feedback, and (2) includes cognitively plausible mechanisms such as means-ends analysis, search, and motivation. Because all models presented are learning-centric and cognitively plausible, they represent attractive alternatives to the traditional symbolic systems that still dominate problem solving. |
|---|
| responsibles | Pélissier |
|---|
| |
|