Snake — Pathfinding studio with smooth rendering

Snake Simulation

Smooth rendering for pathfinding heuristics and endgame play.

Smooth realtime

Test speed 80 ms

Experience overview

You are watching a live snake agent blend classic pathfinding with learned policy choices. The board renders each step while the agent prioritizes safe routes, food collection, and endgame survivability.

Adjust playback speed, board size, and the BFS/Hamilton overlay to see how routing decisions change.
Switch algorithms, toggle curriculum training, and tune reward sliders to change what the agent values.
The end-game parameter (especially the tail space threshold) is often crucial—small tweaks can determine when the snake switches into safer late-game behavior.

Board size 20×20

Show BFS / Hamilton Overlay

Learning

Auto handles the curriculum, save logic, and hyperparameters for you.

Algorithm

Stage preset

Soundtrack

Add MP3 files to the Audio folder to enable the in-game soundtrack.

Enable Late-Game Curriculum

Starting length 3 Starting pattern

Sets the starting length; with the curriculum enabled longer start shapes are used throughout the run.

Episodes0

Avg reward (100)0.0

Best length0

Best length (test)0

Fruit / ep0.0

Greedy Fruit (avg)—

Greedy Reward (avg)—

Gap (Fruit − Greedy)—

Prioritized replay, n-step returns, and dueling heads make DQN stable and sample efficient.

Auto-run

Rounds to run 0 = unlimited auto-run cycles.

Adjustment tempo

Run at least 100 episodes to see the trends.

Average reward Average fruit Greedy Reward Greedy Fruit

Episodes —

Net reward (avg 100) 0.00 (trend +0.00)

Component	Last	Avg 100	Avg 500	Share	Trend

Reward & tuning

Adjust incentives and hyperparameters.

Reward model

Tempo & direction

Step penalty Penalizes every move so the snake keeps routes efficient. 0.010 Turn penalty Adds cost to direction changes to discourage zig-zags. 0.001 Toward fruit bonus Rewards steps that move the head closer to the fruit. 0.030

Away from fruit penalty Punishes moves that increase the distance to the fruit. 0.030

Loops & revisits

Loop penalty Applies extra loss when the snake repeats tight loops. 0.50 Tight loop penalty Punishes revisiting dense spirals and shrinking free space. 1.20 Revisit penalty Penalizes revisiting recent tiles to promote exploration. 0.050

Crashes & stalling

Wall crash Sets the penalty when the snake hits the board edge. 10.0 Self crash Sets the penalty for colliding with the snake's own body. 25.5 Timeout penalty Charges a penalty if no fruit is eaten before the timer expires. 5.0

Trap penalty Punishes creating dead-end pockets around the head. 0.50 Open space bonus Rewards leaving breathable space around the head. 0.05 Dead-end penalty Scales the penalty when the head enters cramped pockets. 0.50

Tail follow bonus Rewards hugging the tail when free space is running low. 0.10 Tail space threshold Tail bonus activates when the free-space ratio drops below this level. 0.150

High-impact rewards

Fruit reward Defines the main reward granted when fruit is eaten. 10.0 Growth bonus Adds a bonus for extending the snake's length. 1.0 Compactness bonus Rewards tighter body packing inside the bounding box. 0.000

BFS / Hamilton Rewards

Matches with the planner award the full bonus; ignoring the helper costs roughly 25% of the slider value so the agent leans on guidance only when confident.

BFS Weight: 0.20
Hamilton Weight: 0.05
Enable BFS / Hamilton

Advanced settings

Shared

Parallel environments Number of simultaneous game instances for experience collection. 1 γ (discount) Discount factor balancing immediate and future rewards. 0.98 LR Learning rate controlling the optimizer step size. 0.0005

DQN family

ε start Initial exploration probability at the start of training. 1.00 ε end Minimum exploration probability after decay completes. 0.12 ε decay (steps) Number of steps before ε reaches the floor value. 80000

Batch Samples drawn per gradient update from replay memory. 128 Replay size Capacity of the experience replay buffer. 50000 Target sync (steps) Frequency of copying weights to the target network. 2000

n-step Horizon length for multi-step return targets. 3 PER α Strength of TD-error prioritisation when sampling replay. 0.60 PER β Bias correction term that anneals prioritised replay. 0.40

Double Q targets Use the online network for action selection while the target net evaluates the value.

Snake Simulation

Experience overview

Learning

Auto-run

Auto adjustments

Training progress

Reward telemetry

Reward & tuning

Tempo & direction

Loops & revisits

Crashes & stalling

High-impact rewards

BFS / Hamilton Rewards

Shared

DQN family

Policy Gradient

Actor-Critic

PPO

Pathfinding primer – overview

AI Auto-Tune data & logging

How the model responses are interpreted

Quickstart

Algorithms and why they work

Reward guide

Tempo & direction

Loops & revisits

Crashes & stalling

High-impact rewards

Reading the reward telemetry

Automatic reward tuning thresholds

Sliders & how they affect training

Environment & playback

Shared hyperparameters

DQN family

Policy-based methods

Diagnostics and common patterns