Zkousel jsem naprogramovat hrajiciho agenta hry SpaceInvaders metodou Deep Q-learning (keon.io/deep-q-learning/)
Teoreticky zaklad a zajimave clanky o deep Q learning jsou na
Vychazel jsem z kodu a videa ‘Deep Q Learning for Video Games – The Math of Intelligence #9’
Za pomoci knihoven
- Tensorflow (https://www.tensorflow.org/),
- Keras (https://keras.io/) a
- simulatoru Gym (https://github.com/openai/gym)
se mi podarilo uhrat skore 455 bodu. Nejlepsi uhrane score je sice cca 5800 (gym.openai.com/envs/SpaceInvaders-v0/), tento algoritmus pouziva ale algorimtus “Asynchronous Actor-Critic Agents (A3C)”:medium.com/emergent-future/simple-reinforcement-learning-with-tensorflow-part-8-asynchronous-actor-critic-agents-a3c-c88f72a5e9f2 , jen co notebook nebude funet jako blazen, zkusim :).
Me zdrojove kody jsou na disku
- c:\Users\honza\Documents\Projekty\20171208_ai_deep-q-learning\
- Nej agent je ulozen v modelu game_agent_model_scoring_235.h5, 2017/12/8 17:00
Jine zajimave
* AI PyGame test prostredi – pygame-learning-environment.readthedocs.io/en/latest/user/games/waterworld.html
* AI Doom test prostredi (nema port pro Win) – github.com/openai/doom-py
* Clanek – arxiv.org/abs/1312.5602
Videa
Zavislosti k installaci