Play games with a checkpoint¶

The screen is the image and the options are the game's buttons. How the game checkpoints were trained, and how they did: Game training.

Watch a checkpoint play¶

examples/atari_live.py and examples/vizdoom_live.py play in a local window:

python examples/atari_live.py --game <Name> --model <checkpoint dir>      # any of the 104 Atari games in ale-py
python examples/vizdoom_live.py --scenario <name> --model <checkpoint dir>

--model reads atari_frames from the checkpoint, so a two-frame model gets {"images": [previous, current]} with no extra flag (--frames 1|2 overrides); --sample draws from the probabilities instead of taking the top action. Trained for 7 minutes on 20,000 auto-labelled frames, a checkpoint plays ViZDoom basic at expert level (mean reward +75.4 against the expert's +75.8 over 50 unseen episodes).

Score a checkpoint on the games suite¶

modal run modal_app.py::games_eval --model <run>/best

It plays, in one go:

• Atari Freeway, Breakout and Galaxian, at atari_eval's settings, against random play and, where expert data exists, the expert. Galaxian has no expert data, so it is compared with random only.
• ViZDoom basic, against the scripted expert, random and always-attack.
• Maze, at 4×4, 6×6 and 8×8 cells: solve rate, and path efficiency against the BFS shortest path.
• Snake, on a 10×10 board: food eaten and steps survived, against a greedy BFS expert and random.
• Classic control from Gymnasium (CartPole, Acrobot, MountainCar, LunarLander), 10 episodes each: episode return and share solved, normalized between random play (0) and a scripted controller (1). A single frame hides velocity, so the screen ghosts the previous frame under the current one.

Maze and Snake are small seeded games in laya/gridgames.py, and the classic-control wrappers are in laya/controlgames.py, so every checkpoint plays the same levels. maze_eval, snake_eval and control_eval compare several checkpoints on one game.