Experiment result

Bootstrap analysis

As we decribe in Section 3.4.1, bootstrap is a method to estimante performance from sampling recorded data. In this analysis, we examine the probability of achievement and accuracy in different voteing threshold. Because we need to observe the analysis to select the suitable voting threshold described in Section 3.3.2. In addition, we also compare the accuracy between voting method and only SWDA be used.

• The analysis results of Subject 1 in three condition data are shown in Figure 4.4.

The results show the equal performance in three condition analysis. In addition, the high enough accuracy with high enogh probability of achievement is presented while setting the threshold to 85%. The comparsion of accuracy of three analysis are shown in the figure right colum.

• The analysis results of Subject 2 in three condition data are shown in Figure 4.5.

The results of subject 2 have the better performance in yes and no condition. The probability of achievement in end condition are lower than the others. However, the accuracy in end condition is similar to yes and no condition whhile the threshold is stted above 80%.

• The analysis results of Subject 3 in three condition data are shown in Figure 4.6.

As the comparsion of accracy show in the figure right column, the accuracy are sig-nificantly improved by using threshod in different percentage of votes. In addition, the reslut of no condition present the best performance.

(a) Yes condition

Figure 4.4: Analysis results in different percentage of votes from Subject 1. Figure a,c and e show the probability of achievement and discriminant accuracy in seven threshold (70, 75, 80, 85, 90, 95, 100 percentage of votes), and the epoch number range from 1 to 10.

The colors in the figure are defined in the color bar which show the degree. Figure b, d and f show the comparison of accuracy of the threshold be set to 85%, 100% and only using

40

(a) Yes condition

Figure 4.5: Analysis results in different percentage of votes from Subject 2. Figure a,c and e show the probability of achievement and discriminant accuracy in seven threshold (70, 75, 80, 85, 90, 95, 100 percentage of votes), and the epoch number range from 1 to 10.

The colors in the figure are defined in the color bar which show the degree. Figure b, d and f show the comparison of accuracy of the threshold be set to 80%, 100% and only using

41

(a) Yes condition

Figure 4.6: Analysis results in different percentage of votes from Subject 3. Figure a,c and e show the probability of achievement and discriminant accuracy in seven threshold (70, 75, 80, 85, 90, 95, 100 percentage of votes), and the epoch number range from 1 to 10.

The colors in the figure are defined in the color bar which show the degree. Figure b, d and f show the comparison of accuracy of the threshold be set to 85%, 100% and only using

42

Online simulation

In this section, we simulate the offline data as a online recorded data. In this procedure, we testing the data trial-by-trial as online recorded until the threshold be reach. Moreover, we evaluate the performance in different threshold setting. Additionally, eighteen datas of each subjects are examined in this procedure.

• The online simulation results of Subject 1 are shown in Figure 4.7. As we expected, the higher the threshold the more data can reach the threshold early. In addition, the result show the good accuracy while the threshold is set above 70%.

• The online simulation results of Subject 2 are shown in Figure 4.8. The growth of the datas number in the 80% threshold which is significantly distinct from the 90%

and 100% threshold. Additionally, the best accuracy is presented while the threshold is set above 90%.

• The online simulation results of Subject 3 are shown in Figure 4.9.

The accuracy are the same by setting the threshod above 70%, and the 83% mean three of the eighteen datas are incorrect.

Bit rate

In this section, we evaluate the performance of each subject by bit rate which decribe in Section 3.4.2,and the maxmum and average information transfer rate are shown in Ta-ble 4.1. The information transfer rate in two analysis procedure are evaluated. Futhermore, we show the two transfer rate in bootstrap procedure. First, the result of classifying only by SWDA. Second, the result of classifying by voting in 80% threshold.

3 4 5 6 7 8 9 10

Figure 4.7: Result of online simulation of Subject 1. The upper figure shows the percentage of data which can achieve threshold under the number of trials. The thresholds are set to 50%, 60%, 70%, 80%, 90% and 100%. The bottom figure shows the discriminant accuracy in the six thresholds.

Figure 4.8: Result of online simulation of Subject 2. The upper figure shows the percentage of data which can achieve threshold under the number of trials. The thresholds are set to 50%, 60%, 70%, 80%, 90% and 100%. The bottom figure shows the discriminant accuracy in the six thresholds.

3 4 5 6 7 8 9 10

Figure 4.9: Result of online simulation of Subject 3. The upper figure shows the percentage of data which can achieve threshold under the number of trials. The thresholds are set to 50%, 60%, 70%, 80%, 90% and 100%. The bottom figure shows the discriminant accuracy in the six thresholds.

Information transfer rate of offline analysis

Subject 1 Subject 2 Subject 3 Bootstrap analysis result Max bit rate/min 7.70 (83.40%) 3.64 (82.10%) 5.74 (76.90%)

(SWDA) Average bit rate/min 2.45 (66.69%) 1.63 (68.01%) 1.65 (64.98%) Bootstrap analysis result Max bit rate/min 7.70 (83.40%) 4.88 (78.40%) 7.09 (97.93%) (80% votes) Average bit rate/min 3.24 (84.15%) 2.75 (89.03%) 2.67 (73.82%) Online simulation (80% votes) Max bit rate/min 5.28 (100.00%) 5.28 (100.00%) 2.26 (100.00%)

Table 4.1: The maxmum bit rate/min and average bit rate/min of three subjects by the dis-crimination only SWDA, voting threashold be set to 80% and testing in online simulation procedure.

4.2 Online

Experiment paradigm

The online experiment paradigm is similar to offline experiment paradigm. However, a decision is only made in the online experiment paradigm, and feedback will be shown on the screen after decision making.

4.2.1 Training

200 250 300 350 400 450 500 550 600 650 700

−4

−3

−2

−1 0 1 2 3 4 5

Three condition averaging data of target

Time [ms]

Amplitude [µv]

Yes condition No condition End condition

Figure 4.10: Three condition averaging data of target.

Three condition averaging data of target. Ecah line represent different condition P300 response.

Before online experiment, subjects have to participate in a training session. The ses-sion is the same as offline experiment. In this seeses-sion, three run of each stimulus (YES, NO, END) are included. Subjects are asked to focus on specific target that is assigned by experimenter before each run. The three condition averaging data are shown as figure 4.10.

Even though the averaging data present the similar P300 response. However, the delay and ampltitude of response are stiil slightly different for three condition. Thus, the three run of

data are used to be training data.