CTC - Car Tycoon Challenge #3 [RESULTS]

[Reaper392]

From what I understand of the types of neural network I studied at university they probably wouldn't be very well suited to this competition. The first problem for using a neural network as a sales model for this competition is that they would need some data to learn from, and because of the difference in the scores that make a 'good' car in the various categories in the different manufacture years I just don't think a neural network would be able to provide an accurate sales model with the amount of data we have from previous competitions. The other problem is that if a neural network spits out strange results they are incredibly difficult to fix.

This is all assuming I haven't missed a very obvious way of applying neural networks to the sales model, or a different type of neural network that would solve these problems

[Frankschtaldt]

Should there be some way of preventing some cars from selling into segments that don't make any sense??
For an example, take my car, a 120kW, 2WD medium sedan with a 4 speed auto that takes almost 10 seconds to get to 100km/h.
It makes no sense at all that something like this would sell to someone looking for a super car yet somehow, 76 people that were looking for a super fast, super expensive super car walked out of a dealership with one of my cars.
It'd be like someone walking out of a Ferrari dealer to go and buy a Mondeo.

Honestly, I don't have an answer but it seems there should be some sort of cut-off that says if a car is too many points behind the market leader for a segment then it doesn't get any sales in that segment.

[xwing]

Especially when there were quite a lot of good SC to choose from. On the other hand, 76 customers is negligible in the grand scheme of things, and you can always imagine someone getting rich quickly, going to try out that Ferrari he/she kept dreaming about the whole life and finding out it isn't really the best choice for daily commute, even if money is not an objection

Or wife/husband stepping in in the last second

[Killrob]

Yes, I don't think that sounds unreasonable... maybe that supercar wasn't quite the way the buyer liked, so another round of driving around a reasonable car it is!

[Slim Jim]

It'd be like someone walking out of a Ferrari dealer to go and buy a Mondeo.

.

Or an Aston Martin Cygnet... just saying. Stranger things have happened.

[nialloftara]

I have achieved consistency, consistently bad but also consistently nailing my optimum price, hey at least I have something. Next challenge maybe I'll stop over engineering my cars and actually pick a target group rather then just building quirky cars.

[xwing]

Idea for the next challenge - how about year 1940? See how many variations of one bodytype can we make?

[strop]

Good lord, making a track day car from a small factory and not going broke... That'll be a challenge for the ages. I don't even think "down force" was a thing back then!

[Frankschtaldt]

Good lord, making a track day car from a small factory and not going broke... That'll be a challenge for the ages. I don't even think "down force" was a thing back then!

LOL, nope. People still didn't really understand how a wing on a plane worked yet let alone how to put one on a car.

[Der Bayer]

Idea for the next challenge - how about year 1940? See how many variations of one bodytype can we make?

Hm, I'd prefer a real test of the new calculations first. But I guess the very early years will come, no matter how many body types there are. It'll be more like a fun challenge then I guess.

[Der Bayer]

Here's some news for the hard-limit-fetishists:

I got an idea how cars can be penalised for not being sporty/prestigious/comfortable, ... enough in a specific target group. I name it the "floating hard limit". It works this way: Each stat is checked if it is lower than a group-specific threshold which directly comes from the weight of the stat in that group. If it is higher, the penalty factor for the overall score is 1 (no penalty). If it is lower, the penalty factor gets a value between 1 and 0, depending on how much too small the value of the stat is. If it is way too low, the factor and thus the overall score will be 0 and the car will not sell in that category. The factors of each stat are multiplied.

The threshold is made that way: For each percent of the weight of the stat, a certain minimum stat value is expected. So higher weighted stats need to be higher than lower weighted stats in order to not get penalised.

So we all get what we want: Cars which are unrealistically bad in one stat for a specific group will not sell. Still, if they are not too shabby, they will only be penalised. And I don't have to balance any hard limits. The thresholds can be scaled with one single coefficient (i.e. the minimum stat value per percent of weight). This is probably necessary because stats are lower in the earlier years than in the later years in general.

Here's how the results of CTC #3 would have looked like with the new penalties:



What do you think? If everything seems to be alright, I will launch the system in CTC #5 (not the current one, that would be unfair to the people who sent in their cars already).

[VicVictory]

Interesting number shifts, but I think I'd have to see how the affect a second set of data (CTC4) before making a judgement on it. Obviously, you'd run the official results for CTC4 as they are submitted now, but it would be interesting to see how these changes would (unofficially) alter the results of the current contest.

[Der Bayer]

I will post the results of CTC4 with the new calculation after the challenge has ended.

[Janekk]

Why wouldn't you put it in CTC #5 ? Even if there are some problems these can always be fixed for CTC #6 and it's not like system is absolutely perfect as it is (if anything why did you make this in first place ? ).

[Der Bayer]

If you say it is shit I won't implement it. If you say it is an improvement compared to before, it's worth a go. I always think that what I do is awesome so I better ask for your opinion. We have the possibility to do a first check with old results, so why not use them and already find some problems?