This software was largely created by AI Vibe Coding
Created by YouMinds
Learn the basics of AI and experiment with a neural network consisting of a single neuron,
optimize parameters and experience how AI algorithms develop in real time.
Perfect for anyone who wants to understand the fundamentals of AI,
from aspiring data scientists, tech enthusiasts to average Joe.
The problem
Below you can see a car on a street. There is a parking lot at the end of the street. The car can drive and
brake using the accelerator pedal. The goal is to park the car perfectly.
Speed: 0 m/s
Distance: 100 m
Time: 0 sec.
Loss: 0
Bang!
Decelerate:Accelerate
Accelerator
Move the accelerator to the right to drive the car and left to brake.
You have 4 seconds to reach the parking lot.
The loss tells you how well you parked.
Let a single neuron AI drive the car
We use a simple artificial neural network consisting of a single neuron with an identity activation
function.
This AI can be described with a simple mathematical formula.
The parameters Weight and Bias
determine the intelligence of the AI.
Output = (Weight X Input) + Bias
Now we apply this AI model to our problem,
we use the distance of the car to the parking lot as input and the output controls the accelerator.
That's all we need to drive the car using AI.
Accelerator=(XDistance)+
Press the “Start AI drive”
button to have the car park using AI. Let's see how well it works.
Change the weight and bias values to improve
the AI.
Train the AI automatically
Training a neural network means finding the weight and bias parameters of all neurons.
In the following table we simply try out different values for weight and bias to find the best one. I.e.
the parameters that minimize loss.
Press the “Create Loss table”
button to calculate all loss values and find the smallest one.
Press “Gradient Decent” and
select a “Start” cell to find a way to the smallest loss as
efficiently as possible.
Note that the simulation runs 10x faster during training.
What is an artificial neural network anyway?
Simply put, an artificial neural network is a mathematical formula that accepts input values and produces
output values,
whereby any number of formulas can be linked and
the output values of one neuron become the input values of the next neuron.
Weights and biases determine the behavior of the formulas or the network.
Finding these values is no different than training the network.
A neuron can therefore be described with a simple expression. The input values x
are each multiplied by a weight w and summed ∑. A bias b
is then added. The result is the network input, which then receives its output value o
from an activation function φ.
Which in our case is an identity function, i.e. y = x.
o = φ (w1 x1 + w2 x2 + ... + wN xN +
b)
o = φ (∑ wi xi + b)
Connect the outputs of one neuron with the inputs of other neurons and you have an artificial neural
network.
Since we only use one neuron in our example and the activation function is omitted, our network can be
formulated simply like this.
Nevertheless, it fulfills all the properties of an artificial neural network. It is trainable and can be
used as a model for solving problems.
o = w x + b
Reinforcement learning
The network shown here uses a special case of reinforced learning for training.
This is often referred to as continuous control in reinforcement learning. Instead of deciding between a
finite set of actions (e.g., accelerate or decelerate), the network produces a continuous range of values.
In reinforcement learning, the neural network can be thought of as an agent because it interacts with the
environment (the car and parking scenario) and learns from it. The values it produces (e.g., throttle
positions) are indeed actions in this continuous action space. The car's motion resulting from these actions
is observed, and based on the outcome (e.g., did the car park correctly?), a reward is given. Over time, the
agent learns to output the values that will maximize the cumulative reward. The reward is the higher the
smaller the error when parking. This error is called the loss.
How does Gradient Decent work?
The method here resembles a form of grid search or manual gradient descent, where you systematically explore
the weight and bias space to find the configuration with the lowest loss. This approach involves testing
multiple combinations of weights and biases, running simulations, and selecting the best-performing set.
A note for experts: Backpropagation cannot be used because it assumes that the loss function based on
the final position of the vehicle is not differentiable with respect to the policy parameters.
Why doesn't Gradient Decent find the best solution?
Gradient Decent follows the next best solution down the valley.
The solution space (loss table) usually consists of many valleys.
Since Gradient Decent cannot see into the next valley and,
in reinforcement learning, only estimates the valley path,
the method gets stuck in a so called local minima.
The global minima would be the best solution that can be found, for example,
using other methods such as Simulated Annealing or,
as in this example, by completely searching the entire solution space using the "Create loss table" button.
Learn how to find valley in machine learning
here.
How was it built
This software was created using Vibe Coding by a Large Language Model LLM / chatbot
and reworked in look & feel.
Some features had to be implemented manually and
corrections and improvements had to be made.
The following Vibe Coding prompts were used on Copilot:
"create a single page html page with embeded javascript and a canvas for 2D drawing. draw a vertical
street
on the canvas and a car on the bottom end of the street. On the top end draw a barrier. Implement a
slider
with range -1 to 1 steps 0.01. when the slider value is above 0 let the car accelerate when the slider
is
below 0 let the car decelerate. when the slider is 0 let the car move with its current speed. when the
car
drives beyond the barrier start over."
"add a field for the car speed based on a street length of 100m. Also add a field for the car distance
to
the barrier. on deceleration do not backup instead stop the car when speed is equal to zero or becomes
negative. add start button that enables the car otherwise it will not start. once it started it
activates
the acceleration and deceleration and a timer starts running. display the timer in seconds. Reset the
timer
upon start over."
"when the car touches the barrier let the barrier show in yellow for 3 second and print the text "bang"
above the barrier for 3 seconds. add a field with a loss function. after 10 seconds measure the loss.
let
the loss be 0 when the car is within a range of 2 m before the barrier and the speed is 0. otherwise let
the
loss be higher Otherwise, let the error increase the further the car is in front of or behind the
barrier
and the greater the speed is after 10 seconds."
"reset the loss on start button press. when the car touches the barrier take the speed as the loss
right
away and do not wait until 10 seconds finished. below the scene add a simulation for a artificial single
neuron. print the formula with input value the weight and a bias and the output. add another button
called
"start AI". When start AI is pressed start the car simulation but let the slider by controlled by the
output
of the neuron and the input to the neuron is the distance of the car to the barrier. Let the user change
the
value for the bias and the weight."
"make shure the simulation stops after 10 seconds and the loss is computed. add another button called
"learn AI". when learn is clicked run the ai simulation using different values for the weight and the
bias
for each run. Use 10 weights in the range from 0 to 0.03 and 10 bias in the range from 0 to 1. add a
table
and use te weights for the columns and the rows for the bias. after each run, when the loss is computed,
write the loss into the corresponding table cell and proceed with the next weight and bias."
"change the bias range in the learn ai. Let the bias run from -1 to 0. Also there is a but, all the
buttons
seem to start the learn ai mode. It must only start with the learn ai button. Assign background color to
the
table cells. use light and colorful backgrounds for low losses and dark and grey colors for high losses.
Also color the cell text so it can be read with contrast."
"learn AI does not iterate the weights and bias nor does it write the loss into the cell."
