Training
A tabular Q-learning agent playing the same engine you do. It sees the distance to the next hazard, the gap ahead, its own velocity and the clearance overhead — and has one choice per tick: hold or release. Its Q-values live in a lookup table, one row per discretized state.
Episodes
1
Steps
0
Avg progress
0.0%
Best progress
0.0%
Clear rate
0%
Avg reward
0.0
Epsilon
1.000
States seen
0
Steps/sec
0
Hyperparameters
Applied live to the running agent — no reset needed. Reset agent restores the table or network to empty but keeps these values.
Learning curve
Avg progressClear rate
Start training to see the curve.
Watch the agent
Runs the greedy policy with no exploration. Before any training it falls back to the hand-written baseline.
Checkpoints
A frozen copy of the agent every 500 episodes, newest 20 kept. Pick one to watch it play, or step through them with Restart after a win.
No checkpoints yet — start training to collect some.