Title: Human Experiments Data from 'Learning Safe Behaviour via Justified Human Preferences and Hypothetical Queries'
Creator: Ilias Kazantzidis (ik3n19@soton.ac.uk)
Date README.txt created: December, 2022
Data DOI:
Subject: Safe agent learning, Safe Reinforcement Learning, Human Preferences, Human-Robot Interaction, Robot training time, Parenting, JPAL, JPAL-HA, Thymio
Funders: UKRI
Recommended citation for the data: 'Learning Safe Behaviour via Justified Human Preferences and Hypothetical Queries'
This dataset supports the publication: 
	AUTHORS: Ilias Kazantzidis, Timothy J. Norman, Yali Du, Christopher T. Freeman
	TITLE: Learning Safe Behaviour via Justified Human Preferences and Hypothetical Queries
	JOURNAL: Autonomous Agents and Multi-Agent Systems
Access Information: Data available under a CC BY 4.0 licence
Language: English
Dates: 
	- Data Collected: July 2022
	- Project Start: November 2021
	- Project End: N/A
	- release date: N/A
Location: University of Southampton, Highfield Campus SO17 1BJ
Methodology: 
	- There were two groups of participants (first_group and second_group)
	- Overall there were 40 participants, but two participants were excluded from the analysis as outliers
	- For group_1, the order of methods was permuted and for group_2, JPAL and JPAL-HA order was permutted too
	- For group_1, the data refers to all three methods (Parenting, JPAL and JPAL-HA), run once, from which the first boxplot (Figure 11) in Section 'User Experiments' was extracted 
	- For group_2, the data refers to the 2nd trial executed with JPAL and JPAL-HA, from which the second boxplot (Figure 12) in Section 'User Experiments' was extracted 'User Experiments'
	- Data was created through timer packages in Python running in a standard laptop, connected via wireless network to a raspberry pi controlling Thymio robot
	- For more information see Section 'User Experiments' from the associated paper
Data Processing: Participants' names were given a index from 01 to 38 to be anonymised
List of file names: 
	- For the first group: <participant_index>_<algorithm>.json
		- <participant_id>: from 01 until 16
		- <algorithm>: one of 'Parenting', 'JPAL' or 'JPAL-HA'
	- For the second group: <participant_index>_<algorithm>_2.json
		- <participant_id>: from 17 until 38
		- <algorithm>: one of'JPAL' or 'JPAL-HA'
		- '_2' refers to the second trial run with JPAL or JPAL-HA
File formats: json
File structure:
	- First layer: First or second group of participants
	- Second layer: Index of participants
	- Third layer: Files of each method run from each participant
	- Within a file there are two lists. The first list corresponds to the name of the variable and the second list to the value of the variable in the corresponding place. All files have the same variables.
Variable list:
	["participant", "algorithm", "overall_activity_time", "agent.P_time", "agent.Pextra_time", "agent.Jp_time", "agent.Janew_time", "agent.P_misses", "agent.Jp_misses", "agent.Janew_misses", "robot_time"]
Code lists:
	"participant": number of each participant
	"algorithm": one of 'Parenting', 'JPAL' or 'JPAL-HA'
	"overall_activity_time": overall training time (in seconds), including thinking of the participant, robot moving, and headers
	"agent.P_time": time of Preference queries
	"agent.Pextra_time": time (in seconds) of additional Preference queries asked at the end of a Hypothetical Actions loop
	"agent.Jp_time": time (in seconds) of Justification over Preferences queries
	"agent.Janew_time": time (in seconds) of Justifications over new (hypothetical) actions
	"agent.P_misses": number of misses in Preference queries
	"agent.Jp_misses": number of misses in Justification over Preferences queries
	"agent.Janew_misses": number of misses in Justification over new (hypothetical) actions
	"robot_time": time (in seconds) the robot spent for each transition
