There are a few things that I'd like to announce and preserve for accountability. Unfortunately, Agents Bar News page isn't ready yet so my blog will have to do. Hopefully the following newsletters are "more professional".

Hey everyone,

Welcome to our first Newsletter. Yes, history is happening now.

In these newsletters we intend to present what we're doing and highlight things that deserve it. We will also try to be brief; nobody has time to read "these days".

Coming this month

Custom Agent

We allowed adding custom environments (more below) and now it's time to allow adding custom agents. Our philosophy is to start with interface design and then allow others to fill the void. Just like LEGO, or Minecraft. Since our Agent entity's APIs are now open GitHub/rl-api-definitions anyone can create their agent and make it work with Agents Bar. One thing left is to enable uploading custom agents which is what we're going to be doing this month.

League

If you haven't heard about the League in context of reinforcement learning then think about a league in the context of sports. It's a way of judging agents based on their interactions with others and/or themselves (in the virtual world cloning is allowed and common). An example of this is creating an AI in a multiplayer game and making these AI learn by playing with each other.

The reason for building League is to enable agents learning by interacting with agents of similar abilities. In the case of a competitive task that's playing against someone who is slightly better. Science aside, such behaviour is somehow intuitive to many. One often gets better when a goal is achievable yet it requires more work than usual. Alternatively, if something is too easy or almost impossible to get then it's either boring or pointless.

Last month

Custom Environment

In a similar tone to the custom agent. We have enabled users to upload their environments and use them in training. Please see RL API Definitions to learn the specification. Currently, environments can be added by passing an url to a public repository with docker image. Unfortunately we aren't Google and thus we have limitations on the size, number of environments and rate of uploading. We will update docs shortly to let you know before you hit them. (Yes, not an ideal order of work; sorry.)

Experiments

This one was exciting. Exciting enough to have its own blog post and LinkedIn reach out. You can read either (both point at the blog post) but in short Experiments allow in-cloud agent/environment communication. Arguably this feature made the Agents Bar a viable option as it significantly reduced the latency. Now, requests are sent within milliseconds. For some this still might be quite slow but this isn't our final word. Stay tuned!

Community outreach

Feeling like you want to spread the word about Agents Bar but something is stopping you? Ok, it's time; we release you. Go ahead and tell everyone about the Agents Bar!

Do you think there's something missing? Something should be done differently? Just want to talk? You also have a green light for this. Let us know! We appreciate any feedback. Write to contact@agents.bar or use the form at https://agents.bar/contact.

Also, intentionally at the end, we are launching a Write for Agents Bar program. This is a paid program for anyone to write about Deep Reinforcement Learning and/or using Agents Bar. It hasn't been widely announced yet as we anticipate some work from our side and there's already a lot on the plate, but hopefully slow rollout will allow us for smooth onboarding.