Randomized Adversarial Imitation Learning for Autonomous Driving

Authors: MyungJae Shin, Joongheon Kim

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we compare the performance between RAIL and baselines. Furthermore, in order to assess the performance gaps between the single-layer and multi-layer policies trained by RAIL, the single-layer and two-layer (i.e., multi-layer) policies was implemented.
Researcher Affiliation Academia Myung Jae Shin , Joongheon Kim Chung-Ang University, Seoul, Republic of Korea {mjshin.cau, joongheon}@gmail.com
Pseudocode Yes Algorithm 1: RAIL
Open Source Code No The paper does not provide any explicit statement or link to its open-source code.
Open Datasets No We make the expert demonstration using PPO with specific action controls. The results present the average of 16 experiment results. In the experiments, the trained weights by BC are used to fast convergence in GAIL and RAIL.
Dataset Splits No The paper mentions average episodes and trajectories but does not specify explicit train/validation/test dataset splits or percentages.
Hardware Specification No The paper does not explicitly describe the hardware used for running experiments.
Software Dependencies No The paper states 'We implemented the RAIL simulator based on Unity' but does not specify version numbers for Unity or other software dependencies.
Experiment Setup Yes Hyperparameters: α step size, N number of sampled directions per iteration, δi and δo Gaussian vectors from zero mean and ν a positive real number standard deviation of the exploration noise, h hidden layer size Initialize : θi, θo from behavior cloning, µ0 = 0 Rn, and P0 = In Rn n