Day-to-day dynamics and FW algorithm
Connections & differences
Traffic equilibrium
[Read More]
Feasible link and path flow set
Convexity and underdetermination
This blog reviews some basic concepts in the traffic assignment. The focus will be on the network structure.
[Read More]
Value iteration process converges to an optimal policy
Why?
In the last policy iteration blog, we prove that starting from an initial pocliy, the iteration process of “evaluation -> greedy improvement -> evaluation -> greedy improvement …” can guarantee an optimal policy. We note that at each step of evaluation, we have to iterate it many times to get...
[Read More]
A copy of online tutorial for Pytorch autograd framework
Self record only
Autograd，自动微分，是整个pytorch框架的核心功能。我一直是只了解原理，但却不知框架是如何实现的，这次看到一个非常好的教程，所以打算把理解总结一下。
[Read More]
Improvimg greedily leads to optimality in policy iteration
Why?
在Dynamic programming里，policy iteration是一个经典的框架，通过反复迭代evaluation和improvement 来得到一个最优的策略。Evaluation得到的值函数，传给improvement过程，而improvement过程根据这个值函数做一个贪婪的policy，返回给evaluation过程。下面看一下为什么贪婪的反馈能够得到一个optimal policy。
[Read More]