Deepmind is saying

"We then use population based training (PBT) to adjust the parameters of the dynamic task generation based on a fitness that aims to improve agents’ general capability."

If I interpret this sentence in my own way,

"Using the evolution algorithm, Give a high score to a task (game) that improves the general capability of the agent, and select and evolve the task (game)"

However, I want to refuse to differentiate between games and agents.

Because both games and agents are algorithms.

If a game is an agent, and an agent is an agent.

By making two agents fight, if the winning agent has general capability, it also selects and evolves the losing agent (game).

Of course, the winning agent is also selected and evolved.

In this case, both the game and the agent are selected and crossed.

The descendants born in that way fight and if the winning side has general capability, choose both the losing side and the winning side..

Cross the two again...

  • 1
    $\begingroup$ I'm sorry, but your post is very confusing. Please try to be more clear. Also, you must ask a specific question. ai.stackexchange.com/help/how-to-ask $\endgroup$ Aug 25 at 17:55
  • $\begingroup$ @AndreGoulart Which part are you confused about? $\endgroup$
    – Dimer
    Aug 25 at 18:06
  • 2
    $\begingroup$ I really want to give an answer, but I can't find a clear question, therefore I can't understand what kind of answer would you expect. Maybe this link could be helpful: ai.stackexchange.com/help/dont-ask. And also this blog post: stackoverflow.blog/2011/01/17/real-questions-have-answers $\endgroup$ Aug 25 at 19:59
  • 2
    $\begingroup$ @AndreGoulart Hello. Thanks for commenting and asking for clarification. I also agree with you. This user is known to ask confusing and/or unclear questions. Given that you already have more than 500 of reputation, you should be able to vote to close this post, so I suggest that you do it. I could do it directly, but I'd like our regular community members to handle these cases independently of the moderators. Once you vote to close the post, it should appear in the review queue for others to vote too. $\endgroup$
    – nbro
    Aug 26 at 11:15
  • 2
    $\begingroup$ @nbro thanks for the tip and specially thanks for allowing newer members (me) to take initiative! Actually I've tried before to flag it and now I just tried to close it. But I got this error message: "This question has an open bounty and cannot be closed" $\endgroup$ Aug 26 at 14:59

The author of the question seems to think that merging the agent and the environment will give them more power. However, there is no apparent gain on trying to merge both models. I found no references to such approach, so it is hard to say it is a good idea. But I can see some reasons why this is probably not a good idea:

1. Fitness Function:

The fitness function is not the same for the agent and the environment.

  • The agents are rewarded by their score in the game.
  • The environments are measured by their ability to train an agent. If they have different roles and different goals, they should use different models (like GANs).

2. Convergence and Volatility:

When training any large model, the learning rate will eventually get very low and the changes get slow and subtle. That is good for the agent, but bad for the environment. The environment should be volatile and change quickly, otherwise the agent will overfit.

3. Transfer Learning:

The author suggests that both (agent and environment) should benefit from sharing the same core. This might be true for when both models do the same task. However, if they do different tasks, then sharing the same core might be harmful:

3.1 Adversarial Training.

In adversarial training, sharing a model might be harmful because it's hard to detect your opponents flaw by thinking just like them.


A General Purpose AI is not created just by solving 2 opposite tasks.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.