TimZaman / dotaclient

distributed RL spaghetti al arabiata
26 stars 7 forks source link

Better Exploration and Training #32

Open Nostrademous opened 5 years ago

Nostrademous commented 5 years ago

This is its own issue regarding agent expectations and the comment you made with respect to win/loss reward and it not mattering currently.

I think that becoming a 'last-hit' master is currently all the agent can hope to accomplish. I say this because there is the following intrinsic rewards:

1) Stay near the center (sure a micro-reward but it still counts) so it encourages the agent not to "explore" through negative reinforcement 2) The bot cannot do anything other than move or attack enemy NPC units or heroes or deny own units. It has no concept of enemy tower, just its own tower and only for purposes of knowing its health but nothing else about it. It is encouraged to last-hit, deny, kill enemy hero, not die. 3) It could potentially learn to help push the creep wave into enemy tower and thus "win" the game but it is discouraged from doing so currently as the reward for killing enemy hero is 3x the reward for winning a game early. (you kill enemy hero twice at +3.0 reward for each kill to win and thus gain the +1.0 reward for win, or you destroy 1 tower which the bot cannot even attack as it doesn't have that unit handle for no-reward other than the +1.0 win reward)

As a result, I think becoming a last-hit champ is all it can aspire to.