OpenAI Gym

OpenAI Gym | “Ant-v1” : Faire marcher une créature à quatre pattes le plus vite possible

Nos agents AI surhumains actuels peuvent percevoir, apprendre de l’expérience, simuler notre monde et orchestrer des méta-solutions (AI forte).

La référence internationale est l’OpenAI Gym.

Tagged: , ,

Comment (1)

  1. Honghu

    Dear distinguished researcher from MontrealAI’s group,
    I am currently working on my master thesis, trying to solve the a robot manipulation tasks through reinforcement learning. However, the code I write myself or from other people don’t perform very well even on openAI classic control scenarios, basically instability issue is the key obstacle, where training reward doesn’t show a steady increase with growing training episodes in DQN. Many endeavors have been made, but didn’t achieve good results, either.
    I see the code implemented from your side gaining really stable performance, however, the code is not available online.
    (1)https://gym.openai.com/evaluations/eval_X07D63uzRka1QoJEAfty7g
    (2)https://gym.openai.com/evaluations/eval_JBCrgRCWSRGSgjrR25jAgg
    (3)https://gym.openai.com/evaluations/eval_HfiRBGmQySwlIiePUxCkw
    I wonder whether I could take a look at this code or at least could I know which reinforcement learning method you are applying so that a good stability as well as a good accumulative reward can be reached? I will be really grateful to that!

Leave a Reply

Your email address will not be published. Required fields are marked *

Pour parler à un être humain

Contactez-nous!