vilvik

BETA

Light
Dark
Auto

Home
Features
Pricing
Blog
Contact us
App

How was your experience?

Leave this field empty

Your feedback helps us improve. Pick a star rating; everything else is optional.

Tap a star to rate

Tell us more (optional)

Name

Only share these if you'd like a reply.

Ask me again later automatically

You can share feedback any time using the Feedback button in the corner of the page.

Thanks for the feedback!

All demos

Share to

Flappy Bird

Live · No sign-in needed

Demo 1, Live Game-Playing Agent

Flappy Bird
Snake
Self-Driving Car
Dino Runner
Breakout
Pong
Gem Collector
Road Crossing

Agent You

Auto-continue between episodes

0 Best 0

Training generation 1…

Connection lost

Click Start to reconnect

Tap to flap

Space / Enter Start, then Space Pause / Resume Esc Reset M Switch mode Space / tap canvas. Start / Flap P Pause / Resume Esc Reset M Switch mode Space Start ↑ ↓ ← → / tap d-pad Steer P Pause / Resume Esc Reset M Switch mode Space Start ← → / tap d-pad Steer P Pause / Resume Esc Reset M Switch mode Space / ↑ Jump ↓ / hold to Duck P Pause / Resume Esc Reset M Switch mode Space Start ← → / tap d-pad Move paddle P Pause / Resume Esc Reset M Switch mode Space Start ↑ ↓ / tap d-pad Move paddle P Pause / Resume Esc Reset M Switch mode Space Start ↑ ↓ ← → / tap d-pad Move P Pause / Resume Esc Reset M Switch mode Space Start ↑ ↓ / tap d-pad Move P Pause / Resume Esc Reset M Switch mode

Training

Status: Idle
Cycle: —
Best score: —
Last cycle score: —
Max score / episode: —
: —

Best score per cycle

About, Live Game-Playing Agent

Evolving a 5→6→1 neural net to decide when to flap.

A small neural network learns to flap through pipes. The Genetic Algorithm (GA) evolves the network's weights between generations.

How it works

The network has 5 inputs (the bird's height, velocity, and distances to the next pipe), one hidden layer of 6 neurons, and 1 output. If the output crosses a threshold, the bird flaps.
Every member of the population plays a Flappy Bird run. Fitness is the time the bird survives plus a small bonus for each pipe it clears.
The top performers are selected as parents. Their weights are recombined and mutated to seed the next generation.
You see the current best policy net play live, and a chart tracks fitness over time.

What to watch

Early generations crash into the first pipe. Within a few generations the bird learns to pace its flaps.
The fitness curve flattens once the policy is solid; spikes after that are the rare lucky runs.

Settings

These settings tune the Agent-mode GA training only. User mode plays the env directly. There's no GA in the loop, so none of these sliders apply to it.

New values apply on the next Start, they're not safe to swap mid-cycle since the population shape is locked once a run begins. Click Reset first if a session is active.

Population size sol_per_pop —

Number of candidate networks per generation. Larger = more diversity, slower per cycle.

Generations / cycle num_generations —

How many GA generations run inside each pygadplus call. Higher = faster convergence, but Pause / Reset latency grows because one call is uninterruptible. Capped at 15 for production safety.

Cycle cap max_cycles —

Stops after this many training cycles. The dashboard shows current/total.

Max steps / episode max_episode_steps —

Per-episode step budget. For Flappy this also caps the visible score (≈ steps / 36 pipes).

Mutation rate mutation_percent_genes —%

% of genes (network weights) randomly perturbed each generation. Higher = more exploration, less convergence.

Number of blocks 12

How many obstacles the grid gets, re-randomized each run. 0 is an open grid.

Traffic density 4 / 5

How busy the lanes are. Higher means more cars and tighter gaps to time.

Playback speed 5 fps

How fast the replay plays, both the frog and the traffic. Lower is slower, so each decision is easier to follow.

See all demos Sign in to build your own

Share to

Flappy Bird

Demo 1, Live Game-Playing Agent

Training

Best score per cycle

Want to run your own?