Directed-Info GAIL: Learning Hierarchical Policies from Unsegmented Demonstrations using Directed Information

Authors: Mohit Sharma, Arjun Sharma, Nicholas Rhinehart, Kris M. Kitani

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present results on both discrete and continuous state-action environments. In both of these settings we show that (1) our method is able to segment out sub-tasks from given expert trajectories, (2) learn sub-task conditioned policies, and (3) learn to combine these sub-task policies in order to achieve the task objective.
Researcher Affiliation Academia Mohit Sharma , Arjun Sharma , Nick Rhinehart, Kris M. Kitani Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {mohits1,arjuns2,nrhineha,kkitani}@cs.cmu.edu
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No A video of our results on Hopper and Walker environments can be seen at https://sites.google.com/view/directedinfo-gail.
Open Datasets Yes we also show experiments on Pendulum, Inverted Pendulum, Hopper and Walker environments, provided in Open AI Gym (Brockman et al., 2016).
Dataset Splits No We used 25 expert trajectories for the Pendulum and Inverted Pendulum tasks and 50 expert trajectories for experiments with the Hopper and Walker environments.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models.
Software Dependencies No We used Adam (Kingma & Ba, 2014) as our optimizer setting an initial learning rate of 3e 4. Further, we used the Proximal Policy Optimization algorithm (Schulman et al., 2017) to train our policy network with ϵ = 0.2.
Experiment Setup Yes Table 3 lists the experiment settings for all of the different environments. We use multi-layer perceptrons for our policy (generator), value, reward (discriminator) and posterior function representations. Each network consisted of 2 hidden layers with 64 units in each layer and Re LU as our non-linearity function. We used Adam (Kingma & Ba, 2014) as our optimizer setting an initial learning rate of 3e 4. Further, we used the Proximal Policy Optimization algorithm (Schulman et al., 2017) to train our policy network with ϵ = 0.2. For the VAE pre-training step we set the VAE learning rate also to 3e 4. For the Gumbel-Softmax distribution we set an initial temperature τ = 5.0. The temperature is annealed using using an exponential decay with the following schedule τ = max(0.1, exp kt), where k = 3e 3 and t is the current epoch.