Learning Programmatically Structured Representations with Perceptor Gradients

Authors: Svetlin Penkov, Subramanian Ramamoorthy

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results confirm that the perceptor gradients algorithm is able to efficiently learn transferable symbolic representations as well as generate new observations according to a semantically meaningful specification.
Researcher Affiliation Collaboration Svetlin Penkov 1, 2, , Subramanian Ramamoorthy 1, 2 1The University of Edinburgh, 2Five AI {sv.penkov, s.ramamoorthy}@ed.ac.uk
Pseudocode Yes Algorithm 1: Perceptor rollout for a single episode
Open Source Code No The paper mentions using third-party packages like 'python-control' and 'python-astar', providing links to them. However, it does not state that the authors' own implementation code for the perceptor gradients algorithm is open-source or publicly available.
Open Datasets No For the cart-pole experiment, the paper states 'we generated a dataset of observations obtained by controlling the cart-pole with the perceptor for 100 episodes.' For Minecraft, the environment is described without mention of a specific public dataset or access information for the generated data. The paper uses OpenAI Gym and a Minecraft-like environment, but doesn't provide specific access to the *data* they used for training.
Dataset Splits No The paper does not explicitly provide specific percentages, sample counts, or citations to predefined splits for training, validation, and testing datasets.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory amounts, or detailed computer specifications used for running its experiments.
Software Dependencies No The paper mentions using 'python-control' and 'python-astar' packages and 'Open AI gym' but does not specify their version numbers or other software dependencies with explicit versioning (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We set the program ρ to implement an LQR. The perceptor ψθ is a convolutional neural network (see A.1)... Q = 103 I4 and R = 1... producing 2 discrete actions required by the Open AI gym cart-pole environment. ... we set the program ρ to be ρ(σ) = a = 1 if Kσ > 0 0 otherwise... We train both perceptors jointly, but keep the learning rate for the pre-trained pose perceptor considerably lower than the one for the wood perceptor.