MSC: Experiment running on all 5 PCs in Lab

Set up all 5 PCs to run experiments now. Will start Experiment on track 2 tonight or tomorrow morning.

Plan to meet Dr James tomorrow, Oct 7, Friday.

Write Paper 2!!

MSC: Torcs Suprisingly Runs on Vista - Finally

Tried running the experiment of Sigmoidal ANN on PC2's native OS, Vista, and surprisingly returned no error. Will duplicate it and run the experiment on PC3 and PC5, which also runs Vista.

MSC: Gen2 ANN Experiment

Today start the experiment with the 2nd generation sigmoidal ANNs.

MSC: The evolved controllers beat me

there were some round, the evolved controller  had almost took over the leading car, but bangged on it and both the cars skidded, but the evolved controller could not recover from the crash and stayed there.

lap7: > 10sec behind

the controllers evolved on track 1 usually started slow, most likely due to the unoptimized automatic transmission, but would later overtake its oponents after one or two laps.

controllers evolved on track 2 were able to generalize on track 1, but could only complete one fifth of track 3.

controllers evolved on track 3 were able to generalize on both track 1 and 2.

controllers evolved on track 1 were not able to generalize on both track 2 and 3.



i lost to the evolved controller on both track 1 and 2, butnot track 3 despite the difficulty controlling the car.


the gear change played a very big impact

MSC: Suggested improvements and new discovery

Received a reply from Dr J yesterday.
Listed below are the suggested improvements and its state:

1. Use round figures to label the x-axis of your graphs (done)
2. Show more in-game screenshots of you SNN-driven car – although it is not of much scientific value here but nice pictures always gets attention
3. Think of plotting the path driven by the best evolved SNNs as they move across the track over the racing period
4. Avoid saying that your evolved controllers were so easily beaten by human and some of the included controllers.
5. Also, your explanation that humans cheat by flooring the accelerator is not convincing since if it is allowed in the game, then it is not cheating. What I meant by cheating is when the game AI controllers break rules such as achieving top speed immediately (as opposed to flooring the accelerator but gradual build-up of speed) or in the case of RTS, being able to see through fog-of-war, then that’s cheating. So I recommend that you amend the way you are currently explaining this point.
6. Please add in a detailed description of your fitness function including a mathematical formula for it. (done)

Discovered something new. In track 1, the controller is able to start by flooring the gas pedal, but it is still slower than the other car because of the gear change strategy. The evolved controller increased gear earlier than the static controller, hence picked up speed slower than it.

Drew track path, but...

Notice that
championship2011server uses a car1-trb1 and the others bot using the same car are:

• tita 3
• berniw 3
• olethros 3
• lliaw 3
• inferno 3
• bt 3

MSC: PC3 keeps crashing, another day to completely compile all results

Run on PC2 is done, but run on PC3 was interrupted by a PC crash at 4am, so I continued it at 10.30pm.
Transferred the results of a run from home PC to PC1, to be compiled together with the rest.

Archived a collection of 3 runs on track 1, the latest, "3_snn_1.csv" is guaranteed to run with the latest settings: RS parameters, best lap time, but there's 2 others, which i'm not quite sure. So I'm sending them home to test run it on home PC.

I compiled the results of the 2 latest completed runs, the one from home, and PC2, into "Track1.xlsx".

As of now, PC1 has completed around 650 generations. Will complete tomorrow.

Found out yesterday from the email Dr J sent us that there's gonna be an international conference in Bangkok. Submission due date is on September 10, 2011.

MSC: My laptop is a snail

The run on laptop is running too slow. It's been a day already and it hasn't even completed 100 generations. Since the run on PC1 was done yesterday, I'm transferring the run on the laptop to continue on PC1.

The run on PC3 on the other hand has stopped, cos the computer crashed, around 2pm, so I continued it around 5pm. Run on PC2 is fine. and finishing today.

UPDATE: So I continued the run from my laptop on PC1, but I discovered that the results don't match, in fact, the results varies so much. One reason 1 observed, is because the settings I had on my laptop for the race is 1 lap! but how come?! Oh I remember, I duplicated the copy I had in my HDD, which probably has a setting of only 1 lap. Hence the flawed results. Hence I had taken to start a new run on PC1. What a waste of time.

MSC: table to display best lap time

Table structure to display the best lap times of each run for all 3 tracks.

Run	Track 1	Track 2	Track 3
1	...s	...s	...s
2	...s	...s	...s
3	...s	...s	...s
4	...s	...s	...s
5	...s	...s	...s
6	...s	...s	...s
7	...s	...s	...s
8	...s	...s	...s
9	...s	...s	...s
10	...s	...s	...s
	mean +/- sd	mean +/- sd	mean +/- sd

MSC: Algorithm for best lap time

if(cs.getLastLapTime() > 0.0)
{
    if(bestLapTime == 0.0)
        bestLapTime = cs.getLastLapTime();
    else if(cs.getLastLapTime() < bestLapTime)
        bestLapTime = cs.getLastLapTime();
}

Wasted 4 days running redundant experiments

Discovered that I had wasted 4 days running 2 experimental runs on track 2!! Gosh. So I initiated another run, but this time mad sure it started in generation 0 and the track is track 1.

Will go to school today to start more runs on track 1 in lab. Now, I need some sleep!

SimpleDriver Best Lap Times

A-Speedway: lap1: 01:07:13    Best:01:02:68
CG speedway number 1: lap1: 01:32:96    Best:01:28:65
Street 1: Best:03:06:00

Graphs are on the way

Mingled with the data today.
Wrote a VBA macro to organize the fitness (did not organize the lap time yet. the lap time will be displayed in tables).

Extracted out the highest fitness of each generation in all 10 runs for track 2 and 3, and copied the fitness into a separate sheet, where only all the highest fitness of all 10 runs are displayed, along with a graph. In that separate sheet, named "Overall", included the mean and st deviation of the highest fitness for all generations.
Will see what I will do with it later. Putting them in tables or displaying another graph? or using them to display the mean of all 10 runs of the best individual?

That is 1 solution.

MSC:

Formatted paper to mirror IEEE format requirement, for the paper due Aug 15.

Track 2
2 completed runs on my lab PC, and 1 ongoing run.
2 completed runs on the other PC on the far right, and 1 ongoing run.
2 completed runs on the PC on the right.
2 completed runs on home's PC, and 1 ongoing run.

Total: 11 runs

MSC: Update - July 16, 2011

Update 11/7/11
checked again. fitness score still differ.
- another observation is that - all the fitness scores are the same in generation 0,
- but after allowing it to evolve 1 or 2 times, and checked the fitness score of the eg. 2nd generation,
- the fitness score varied again.
started testing the inputs and then will test the number of spikes throughout the run. printing them into a file

Observations:
found out that the inputs of the 3rd/5th generation differ from the new run with the chromo from the run before. It was the same when the cars are still in "get-set" period, but differed after starting to move, which is most likely caused by differing outputs.
saved the source into pendrive to bring home to investigate, while setup the comps in lab to run experiments with v & u being reset every timestep.
no efforts in write up. cannot be like this! must write!!!

Update 14/7/11
tested more vigorously. this time tested for the outputs. found out that the outputs are not the same. causing the spikes to be different.
hence there are only 2 causes.. it's either the inputs or the weights being different.
but the inputs are the same... hence it should be the weight.. then realized that the weights may have longer trailing floating points.
unlike those that are printed into file which only has 6 trailing max. hence i concluded that the printed weights have less precision compared to the weights stored in the variables.
finally found out why the scores are different.
went home running the experiment with the values of v & u being reset at every timestep.

Update 16/7/11
still running experiments with v & u being reset at every timestep. Spent last night here in lab. it's 8pm now. going home to write!
Lord bless and inspire me. hehe. thanks!!!