you said once that openAI did increase learning scalability to even greater levels. Previously the agent was able to learn 180 years a day via self-play. What is the new learning rate after IT is over?

try another RL scheme is using a modificated SARSA RL structure. We propose the next loop (Feeling, feeling Reward, sensor perception, Action, new feeling, general Reward, New State or perception) feeling - an state of specific internal parameters, inner emotions.

Your AI needs to have a sense of "risk taking" to be able to beat top tier teams. "Safe plays" is not that rewarding in Dota. Item like Divine Rapier is one thing i want to see the AI utilize next year (if there is any next year).