Implementations for the courses, publications and things that can explode anytime...
This page contains some of the programs created during my MSc studies. Some of them are part of the courses and some are the implementations used in my publications. Feel free to play around and please give me a feedback if you feel like doing so. :)
- Road pricing approach for Pigou's example
- Simple Q-learning
- Modified minority game for traffic assignment
Pigou's example is about selfishnes in route choice: given two roads, a main road with the travel time depending on the number of drivers in it and a side road with fixed high travel time, all drivers will opt for the variable-cost road. But when they do that, the number of drivers in that road becomes so high that the travel time is equal to the side road's.
This phenomena makes it useless to have a main and side road. To solve it, a road pricing approach is proposed: drivers must pay credits to use the main road. This way, hasty drivers will use it and economic drivers will prefer the side road. This way, travel times are reduced for hasty drivers while economic drivers save money.
Simple Q-Learning (in portuguese)
This is a Reinforcement Learning experiment. Without prior knowledge, an agent must learn a behavior by interacting with the environment. The agent receives a reward signal by tring a given action in a given environment state. By exploring the state-action space in a reasonable number of times, the agent learns the path to its goal.
This experiiment implements the modified Minority Game algorithm for traffic assignment proposed in the paper: Road Traffic Optimisation Using an Evolutionary Game written by Syed Md. Galib and Irene Moser.
Agents predict occupancy on the roads they can use on their trip. They have a set of predictors for each link and "believe" the prediction made by the best scored predictor. Then they choose the next link of their trip as the one with the least predicted occupancy. This is repeated until they reach their destinations. Then travel times are calculated and the predictors' score are updated.