Human Preferences for Reward Learning in a 10x10 Grid-World Game

A dataset from Texas Data Repository Harvested Dataverse, created by William Bradley Knox and last updated on March 18, 2024. It consists of human preference judgments over trajectory segments in a custom grid-based game, collected to study reward function learning for reinforcement learning agents. The data was gathered from 143 subjects recruited via Amazon Mechanical Turk.

Use Cases

• Benchmarking preference-based reward learning algorithms based on the provided trajectory segment comparisons.
• Studying human decision-making trade-offs in grid-world tasks based on features like obstacle avoidance and coin collection.
• Training models to infer reward functions that balance multiple objectives, as suggested by the game's design.

Strengths

• Includes a detailed data report file (A_data_report.pdf) documenting collection methodology and content.
• Accompanied by the game code used for the study, allowing for replication and extension.
• Data was collected from 143 human subjects, providing a basis for studying human preference patterns.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset's scale and specific file formats are not described in the provided metadata.

Provenance

Source

Texas Data Repository Harvested Dataverse

Collection Method

Human preferences collected via Amazon Mechanical Turk from subjects shown pairs of game trajectories.

Time Range

null

Freshness

Last updated 2024-03-18 10:23:31; freshness should be verified.

Geography

null