Master Thesis

455 views
415 views

Published on

For future interplanetary space missions, there is a need for an autonomous system that prevents astronauts from potentially hazardous situations. One way in which such a system can help the astronauts, is by predicting their performance. This thesis presents the construction of such a system based on Cognitive task load (CTL) and Emotional state (ES) using Bayesian networks. The network is trained and tested with four different datasets.
The first two datasets concern CTL and performance of naval operators working with an adaptive user interface in a lab-setting and naval operators working on a high-tech sailing ship. Especially the ship dataset generalizes very good. The third dataset concerns the ES and performance of participants that are playing a multiplayer first-person shooter game. The last dataset concerns CTL, ES and performance of participants during a three user learning task. The best networks provided respectively a performance estimation of 84.8%, 74.2%, 60.8% and 67.0% while the chance level was 50%. These results indicate that this approach is prommising. Further research is needed to increase the performance such that the method can be used during training and real missions.

Published in: Education, Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
455
On SlideShare
0
From Embeds
0
Number of Embeds
6
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Master Thesis

  1. 1. Performance prediction: preliminary research for manned missions to Mars Stefan kennedie [kennedie[at]gmail.com]
  2. 2. Content <ul><li>Manned space flight </li></ul><ul><li>My assignment </li></ul><ul><li>Research question </li></ul><ul><li>Experiments </li></ul><ul><li>Conclusion </li></ul>
  3. 3. Manned space flight <ul><li>Past: Moon </li></ul><ul><ul><li>Transfer time: 3-5 days </li></ul></ul><ul><li>Currently: International Space Station in Low Earth Orbit </li></ul><ul><ul><li>Transfer time: 1-2 days </li></ul></ul><ul><li>Future: Mars </li></ul><ul><ul><li>Distance: 55-401 million km </li></ul></ul><ul><ul><li>Transfer time: 250 days </li></ul></ul><ul><ul><li>Communication delay: 6-44 minutes </li></ul></ul><ul><ul><li>Extreme isolation </li></ul></ul><ul><ul><li>Extreme social monotony </li></ul></ul><ul><ul><li>No visual link to the Earth </li></ul></ul><ul><ul><li>No possibilities for evacuation </li></ul></ul>Manned space flight
  4. 4. MARS-500 <ul><li>Simulated journey to Mars </li></ul><ul><ul><li>“ Are human beings able to realize a Martian flight, taken all limitations and restrictions into account” </li></ul></ul><ul><ul><li>Institute of Biomedical Problems (IBMP), Moscow, Russia </li></ul></ul><ul><ul><li>105 days experiment (April – July 2009) </li></ul></ul><ul><ul><li>520 days experiment (scheduled start: 2010) </li></ul></ul>Manned space flight
  5. 5. MECA consortium <ul><li>Mission Execution Crew Assistant </li></ul><ul><li>Martian crew has to rely on an artificial life system during their voyage to Mars in order to survive </li></ul><ul><li>“ supports humans and machines to act in a distributed, autonomous but cooperative way” </li></ul><ul><li>Predicting performance </li></ul>Manned space flight
  6. 6. Assignment
  7. 7. Assignment <ul><li>Predict performance (low or high) with </li></ul><ul><ul><li>Cognitive task load </li></ul></ul><ul><ul><li>Emotional state </li></ul></ul><ul><li>Use a Artificial Intelligence technique that </li></ul><ul><ul><li>Is transparent and easy to understand for the user </li></ul></ul><ul><ul><li>Can be modular </li></ul></ul><ul><ul><li>Can be different for each astronaut </li></ul></ul><ul><ul><li>Can be adapted while astronauts are in deep space </li></ul></ul><ul><li>Bayesian network </li></ul>Assignment
  8. 8. Measurements <ul><li>Cognitive Task load </li></ul><ul><li>Level of Information Processing (LIP) </li></ul><ul><ul><li>Skill-based level: automatic processing of information </li></ul></ul><ul><ul><li>Rule-based level: information triggers routine solutions </li></ul></ul><ul><ul><li>Knowledge-based level: information is analyzed, solutions are planned </li></ul></ul><ul><li>Task set switches (TSS) </li></ul><ul><ul><li>Attention shifts to different tasks with different goals </li></ul></ul><ul><li>(percentage) Time Occupied (TOC) </li></ul><ul><ul><li>Should not be >70-80% </li></ul></ul>Assignment
  9. 9. Measurements <ul><li>General problem regions </li></ul><ul><ul><li>Underload </li></ul></ul><ul><ul><ul><li>All: low </li></ul></ul></ul><ul><ul><li>Overload </li></ul></ul><ul><ul><ul><li>All: high </li></ul></ul></ul><ul><ul><li>Vigilance </li></ul></ul><ul><ul><ul><li>Continuously monitoring a process without acting </li></ul></ul></ul><ul><ul><ul><li>LIP: low </li></ul></ul></ul><ul><ul><ul><li>TSS: low </li></ul></ul></ul><ul><ul><ul><li>TOC: high </li></ul></ul></ul><ul><ul><li>Cognitive lock-up </li></ul></ul><ul><ul><ul><li>Focus on a task and reluctance to switch to another task </li></ul></ul></ul><ul><ul><ul><li>LIP: all </li></ul></ul></ul><ul><ul><ul><li>TSS: high </li></ul></ul></ul><ul><ul><ul><li>TOC: high </li></ul></ul></ul>Assignment LIP = Level of Information Processing TSS = Task set switches TOC = Time Occupied
  10. 10. Measurements <ul><li>Emotional state </li></ul><ul><li>Valence (pleasure) </li></ul><ul><ul><li>Positive versus negative </li></ul></ul><ul><li>Arousal (excitement) </li></ul><ul><ul><li>High versus low </li></ul></ul><ul><li>Circumplex model of affect </li></ul>Assignment
  11. 11. Assignment
  12. 12. Measurements <ul><li>Self Assessment Manikin (SAM) </li></ul>Assignment
  13. 13. Research questions <ul><li>Is it possible to predict performance based on Cognitive task load and Emotion using a Bayesian network? </li></ul><ul><li>Is it possible to predict performance based on Cognitive task load? </li></ul><ul><li>Is it possible to predict performance based on Emotional state? </li></ul>Research Questions
  14. 14. Research questions <ul><li>Does the performance of the Bayesian network increase when Emotion and Cognitive task load are used instead of only Emotional state? </li></ul><ul><li>Does the performance of the Bayesian network increase when Emotional state and Cognitive task load are used instead of only Cognitive task load? </li></ul>Research Questions
  15. 15. Experiments <ul><li>Generating datasets with: </li></ul><ul><li>Performance and Cognitive task load </li></ul><ul><ul><li>Marc Grootjen’s naval operators experiment (Lab & Ship) </li></ul></ul><ul><ul><li>Research sub question 1 </li></ul></ul><ul><li>Performance and Emotion </li></ul><ul><ul><li>Paul Merkx’ First person shooter emotion experiment </li></ul></ul><ul><ul><li>Research sub question 2 </li></ul></ul><ul><li>Performance, Cognitive task load and emotion </li></ul><ul><ul><li>MECA’s Collaborative trainer experiment </li></ul></ul><ul><ul><li>Main research question, sub question 3 and 4 </li></ul></ul>Experiments
  16. 16. Cognitive Task Load <ul><li>Lab experiment </li></ul><ul><li>12 students had to deal with alarms during platform supervision, damage control and navigation tasks </li></ul><ul><li>1407 cases with data for LIP, TSS, TOC and Performance </li></ul><ul><ul><li>60 sec with an overlap of 50 sec </li></ul></ul><ul><li>All 8 of 8 networks performed significantly better than chance level (50%) </li></ul><ul><li>Performance of best performing network: 84.8% </li></ul>Experiments
  17. 17. Cognitive Task Load <ul><li>Ship experiment </li></ul><ul><li>24 participants performed three scenario’s on sailing ships </li></ul><ul><li>3478 cases with data for LIP, TSS, TOC and Performance </li></ul><ul><ul><li>60 sec with an overlap of 50 sec </li></ul></ul><ul><li>All 8 of 8 networks performed significantly better than chance level </li></ul><ul><li>Performance of best performing network: 74.2% </li></ul>Experiments
  18. 18. Gognitive Task Load <ul><li>Generizability </li></ul><ul><li>Trained with Ship dataset, tested with Lab dataset </li></ul><ul><ul><li>Max. network performance: 82.6% instead of 74.2% (+8.4%) </li></ul></ul><ul><li>Trained with Lab dataset, tested with Ship dataset </li></ul><ul><ul><li>Max. network performance: 69.5% instead of 84.8% (-15.3%) </li></ul></ul><ul><li>Networks trained with real-world data generalize very nice. </li></ul>Experiments +8.4% Difference 82.6% Lab 74.2% Ship Test set Lab Ship Train set -15.3% +8.4% Difference 84.8% 82.6% Lab 69.5% 74.2% Ship Test set Lab Ship Train set
  19. 19. Emotional state <ul><li>2 games of 20 min 2-on-2 capture the flag in Unreal Tournament 2003 </li></ul><ul><ul><li>Webcam </li></ul></ul><ul><ul><li>Microphone </li></ul></ul><ul><ul><li>Screen captures (once every second) </li></ul></ul><ul><li>Each player annotated their own valence & arousal on a 1-100 scale for every 10 seconds </li></ul>Experiments
  20. 20. Emotional state <ul><li>Performance was determined on in-game events </li></ul><ul><li>467 cases with data for Valence, Arousal and Performance </li></ul><ul><li>2 of 24 networks performed significantly better than chance level </li></ul><ul><li>Performance of best performing network: 60.8% </li></ul>Experiments
  21. 21. Cognitive Task Load & Emotional state <ul><li>Collaborative trainer </li></ul><ul><li>24 participants performed 7 COLT sessions </li></ul><ul><li>Performance was determined using # questions answered correctly and clicking behavior in the learning environment </li></ul><ul><li>306 cases with LIP, TSS, TOC, Valence, Arousal and Performance </li></ul>Experiments
  22. 22. Cognitive Task Load & Emotional state <ul><li>Networks trained with CTL and Performance </li></ul><ul><ul><li>1 of 9 networks performed significantly better than chance level </li></ul></ul><ul><ul><li>Performance of best performing network: 59.2% </li></ul></ul><ul><li>Networks trained with ES and Performance </li></ul><ul><ul><li>0 of 9 networks performed significantly better than chance level </li></ul></ul><ul><ul><li>Performance of best performing network: 57.8% </li></ul></ul>Experiments
  23. 23. Cognitive Task Load & Emotional state <ul><li>Networks trained with complete dataset </li></ul><ul><ul><li>4 of 9 networks performed significantly better than chance level </li></ul></ul><ul><ul><li>Performance of best performing network: 67.0% </li></ul></ul><ul><ul><li>Significant higher performance than best network trained with only CTL data </li></ul></ul><ul><ul><li>Significant higher performance than best network trained with only Emotion data </li></ul></ul>Experiments
  24. 24. Conclusions <ul><li>It is possible to predict performance based on Cognitive task load and Emotion using a Bayesian network </li></ul><ul><li>It is possible to predict performance based on Cognitive task load </li></ul><ul><li>It is possible to predict performance based on Emotion </li></ul>Conclusions
  25. 25. Conclusions <ul><li>The performance of the Bayesian network does increase when Emotion and Cognitive task load are used instead of only Emotion </li></ul><ul><li>The performance of the Bayesian network does increase when Emotion and Cognitive task load are used instead of only Cognitive task load </li></ul>Conclusions
  26. 26. <ul><li>Bayesian networks can be used to predict performance using Cognitive task load and Emotional state, however, more research is needed for applying them in the space domain. </li></ul>

×