Analysis and Visualization of Behavioral Network Science Experiments Francisco J. Gutierrez-Villarreal Computer Science Department, Hartnell College, Salinas, CA 93908 David L. Alderson, Ph.D, Emily Craparo, Ph.D, Operations Research Department, NPS Thomas W. Otani, Ph.D, Computer Science Department, NPS Data Processing & Analysis Overview This project involved analyzing the results of a series of behavioral experiments conducted by researchers at the Naval Postgraduate School (NPS) and at the University of California, Santa Barbara (UCSB). The experiment involved individuals within a community deciding if and when to evacuate from a pending natural disaster. The primary objective of this project was to understand the way in which individual decision makers use and share information, and how this information leads to collective action of the group as a whole. Raw data was processed using the Python computer language in order to obtain information about the influence of personal networks on individual and group behavior. Matplotlib, a Python library, was used to graphically represent subject behaviors and experiment outcomes. Experiment: 18 May 2012, UCSB • 50 Players • 47 Games (1 min each) Analysis (Continued) Log Files, 1 per game (all players) 21 21 21 21 18 18 14 17 18 03 18 48 38 09 32 14 14 43 39 35 18 06 23 18 22 14 27 33 14 02 09 17 17 43 43 09 TAB_SWITCH NODE_PRESS NODE_RELEASE TAB_SWITCH NODE_PRESS NODE_RELEASE TAB_SWITCH TAB_SWITCH NODE_PRESS TAB_SWITCH NODE_RELEASE TAB_SWITCH TAB_SWITCH TAB_SWITCH TAB_SWITCH NODE_PRESS NODE_RELEASE TAB_SWITCH TAB_SWITCH TAB_SWITCH NODE_PRESS TAB_SWITCH TAB_SWITCH NODE_RELEASE TAB_SWITCH NODE_PRESS TAB_SWITCH TAB_SWITCH NODE_RELEASE TAB_SWITCH NODE_PRESS NODE_PRESS NODE_RELEASE NODE_PRESS NODE_RELEASE NODE_RELEASE Social Information 28 28 Disaster Information 43 43 Social Information Social Information 43 Social Information 43 Social Information Social Information Social Information Social Information 25 25 Social Information Social Information Social Information 29 Social Information Social Information 29 Social Information 02 Social Information Social Information 02 Social Information 04 09 09 32 32 04 Color Map that shows each player’s payoff (score) for each game. Rows show players’ scores throughout games. Columns show players’ scores for a single game. Vertical color patterns indicate similar player behaviors in each game. 0 113 185 1669 1747 1842 1877 2197 2201 2324 2505 2536 2598 2734 2768 2846 2950 2998 3004 3005 3074 3147 3168 3242 3441 3442 3447 3489 3534 3606 3658 3698 3769 3833 3897 3922 Histogram shows click frequency distribution (measured in clicks per second) for all players in all games Graphs show the disaster strike probability line (in red), cumulative evacuations as time goes by (blue line), shelter capacity for each game ( black horizontal line), and the point after which a disaster can begin to strike (vertical purple dotted line). One graph of this type was generated for each game. Histogram shows the distribution of total player scores for all games. (Max Score: 4700) Disaster Tab Red Zone is possible disaster strike range Disaster Likelihood Lowest Score Player 11: 1940 points Average: 2344 points Payoff (Max=100) Evacuation Button Social Tab If a player clicks on a neighbor who is in shelter, the bed number occupied by this neighbor is revealed (and stays revealed for the duration of the scenario). The shelter condition is displayed. The number shown is the highest bed number among those occupied by the neighbors player revealed. H Color Maps that show the percentage of time each player spent on each tab for all games. Most players disregarded the social network tab and instead focused on the disaster probability tab. This can be seen in the following graphs. The graph on the left shows the payoff distributions for all possible number of neighbor counts for all games. There is no clear correlation between neighbor counts and performance. The graph on the right shows the clicking activity for each player for each game and lists players by their performance (highest rank first). As the graph shows, there were players who did well and were not very active, and others who were active that did poorly. Since clicking activity is strongly associated with neighbor node clicks, the graphs indicate that group behavior did not heavily influence individual decision makers. Next Steps These results are guiding the development of the next round of experiments, to be held at UCSB in October 2012. Acknowledgments Game Progress Bar Highest Score Player 10: 2590 points Conclusions Graphs show the disaster probability values that a single player saw for each game (probability values are shown vertically for each game). If a player evacuated in a game, a green or red dot is shown. Green dots indicate that a player made the correct evacuation decision (the disaster hit in that game). Red dots indicate that the player evacuated, but the disaster did not strike. One graph of this type was generated for each player. Player rank as determined by total game scores is shown next to the player’s ID number. I would like to thank my mentors, Professors Alderson, Otani, and Craparo from NPS, for sharing their knowledge and experience with me, and for providing patient guidance throughout the project. I would also like to thank Kelly Locke, Andy Newton, Professor Joe Welch, and Pat McNeil from Hartnell College for making this internship possible. Finally, I would like to thank Alison Kerr and Casandra Martin, NPS internship coordinators, for all of their hard work in making the internship program run smoothly, and for helping me navigate through all of the opportunities available at NPS. This internship was funded by Strengthening Transfer Pathways (STP) Title V Grant Office of Naval Research (ONR) Multiple University Research Initiative (MURI) on “Next-Generation Network Science” 2008-2013 For further information Francisco J. Gutierrez-Villarreal franciscogutierrez@student.hartnell.edu David Alderson, Ph.D. dlalders@nps.edu