Two Stanford Symbolic Systems Forum Presentations

Capturing Human Driver Behavior with Generative Adversarial Imitation Learning (Alex Kuefler, advisor: Mykel Kochenderfer, Aeronautics and Astronautics)
The ability to accurately predict and simulate human driving behavior is critical for the development of intelligent transportation systems. Traditional modeling methods have employed supervised machine learning. But due to the long-term, sequential nature of human driving, these models compound mistakes over time - leading to unrealistic highway simulations that include frequent collisions and off-road driving. Using Generative Adversarial Imitation Learning, an algorithm for training deep neural networks to reproduce expert-like behavior based on only a small sample of demonstrations, we overcome this problem of "compounding errors" in realistic highway simulations. Our learned policy network both reproduces emergent behavior of human drivers, such as lane change rate, while maintaining realistic control over long time horizons.

Assessing Resource-Rational Analysis Approaches (Brendan Fleig-Goldstein, advisor: Thomas Icard, Philosophy)
Humans seem to be imperfectly rational in demonstrable ways.  One explanation for this fact is that our brains do not have enough computational resources to perform ideal cognitive calculations in every circumstance.  We might think that our brains perform optimally given the computational resource limitations imposed on them.  If we make this assumption, then we can think about the types of algorithms that the mind might use by thinking about algorithms that make an optimal trade-off between cost and accuracy.  By comparing the performance of such algorithms to human performance, the hope is that we can gain insight into the types of computational resource constraints our minds are subject to, and thereby learn about the functional architectures of our minds. In my talk, I will discuss various ways this approach might be successful and potential difficulties.

Room 126