The main problem with using dropout in reinforcement learning is that the policy network is used twice. Once when collecting rollouts and again during the gradient update. Each time a new dropout mask is randomly generated, meaning the network behaves slightly differently even

Culprit of AI film festival Suhas Sumukh

GRPO vs PPO in the same env.

anyone in delhi?

Delhi.

GRPO merged! find PR here: github.com/Metta-AI/metta…

I am live, join from here: youtube.com/live/p6oqsrAZG…

Revisiting this paper.

Monte-Carlo learning is purely trial and error learning.

Our brain is off-policy learner with on-policy corrections. 1. When we go to sleep we replays experiences during sleep or rest (like the “replay buffer” in RL). That’s an off-policy mechanism. 2. Learning from observation: We can learn by watching others (imitation learning).

built a UI for live training of an agent in gridworld env with Q-learning, you can see how agent is training and how Q score of a state is being updated. During the training you can tweak parameters like learning rate, exploration (epsilon), discount factor and see how training

Wrote a simple article on Q-learning Algorithm. I will also create a video on it :) plume-robin-b8f.notion.site/Q-learning-29b…

RL Madness x.com/i/broadcasts/1…

Congratulations to Women’s Team India for winning the Cricket World Cup. I really enjoyed watching the match. This cup will inspire the whole next generation not just girls but also boys to follow the passion irrespective of any circumstances. It is a great cricket memory to