blog.jeremarc.com/micro_posts/...

👆🏻 RL agent angry about sparse rewards

I just read "Welcome to the Era of Experience", advocating for agents grounded from interacting with the real world and getting their rewards from an "open-world" instead of simulation+simple reward.

I open bluesky

You propose agents working out for you. Someone is preparing for a grant 🤔

An extension could be for multi-agent systems, although way more complex to include the safety constraints (no collision)

arxiv.org/abs/2503.19173

And a great blog post David wrote here: www.davidsilver.uk/wp-content/u...

You can find the paper here: ojs.aaai.org/index.php/AI...

...For every step Agent 1 makes, you reserve 2 cells in the shared table. For every step Agent 2 makes you reserve at least 1 cell.

What if some agents are slower/faster, how would the reservation table look like?
You discretize time based on their relative speed.
Imagine agent 1 4x as fast as agent 2. The shared table has cells of unit 1. Agent 1 table has cells of unit 2 and Agent 2 table cells of unit 0.5...

The idea is explained in this figure.
Agents reserve their shortest path in sequence, and a new Space-Time A* approach is used to find the shortest non-reserved path.