ALA-Cogment Challenge: Announcing the Results!

In January, we were pleased to announce the ALA-Cogment Challenge at the ALA (Adaptive and Learning Agents) Workshop at AAMAS 2022, part of AIR’s interest in supporting the next generation of reinforcement learning research.

We are excited to announce the results! Please see the attached video presentation, which was aired during the award ceremony of the ALA Workshop.

Announcing the Results!

AI Redefined (AIR) is the company that built Cogment, a platform for helping humans and AI agents to collaborate via intelligence ecosystems. Cogment is a great resource for researchers and developers who want to use multi-agent reinforcement learning or human-in-the-loop learning in their projects and applications. That’s why we are proud to work with the ALA Workshop to sponsor the ALA-Cogment Challenge.

When we announced the Challenge back in January, we invited ALA authors to use Cogment’s platform, which is open-source, in their ALA submissions. Cogment allows humans, AI agents, aggregates of AI agents, and non-AI agents to learn together across real-world and simulated environments in real time. Cogment’s implementation swapping can also enable the benchmarking of different actor configurations. Cogment’s versatility makes it a great match for the kinds of projects fostered by the ALA Workshop.

We are happy to announce that the ALA-Cogment Challenge will be giving C$5,000 grants to two teams of authors of ALA Workshop papers.

Introducing grant recipient paper: “Multi-teacher curriculum design for sparse reward environments”

First, authors Chaitanya Kharyal, Tanmay K. Sinha, and Matthew E. Taylor will receive a grant in recognition of their work on their paper “Multi-Teacher curriculum design for sparse reward environments.” This work proposes a novel approach for faster training of goal-conditioned RL agents in sparse reward environments. While most of the existing approaches train only a single agent for exploring better in spare reward environments, this work uses a curriculum set by another agent (teacher) for training the goal-conditioned (student) agent. They improve upon the existing state-of-the-art by proposing a multi-teacher approach. The need for training multiple teachers and student agents simultaneously and completely from scratch made the authors look for frameworks that can handle this in an efficient way, and they found Cogment to be a perfect solution for their needs.

The authors noted that: “The modularity that Cogment-verse provides enabled us to easily and quickly modify it according to our needs for fast experimentation. We would like to continue working on Cogment and contribute to its open source community because we believe that it’s a promising tool for future research in this area.”

Introducing grant recipient paper: “Analyzing and Overcoming Degradation in Warm-Start Off-Policy Reinforcement Learning”

Next, Benjamin Wexler, Elad Sarafian, and Sarit Kraus will receive a grant for their work on their paper “Analyzing and Overcoming Degradation in Warm-Start Off-Policy Reinforcement Learning.” This work attempts to identify the root causes of performance degradation in the pre-trained agents when they start training in an environment using RL algorithms. They also demonstrate that it’s a common phenomenon in offline RL and online RL methods.

Here’s what this team of authors had to say about Cogment: “Cogment is a great resource for experimenting with human-in-the-loop experiments. It makes it easy to run trials across different environments with various common RL algorithms and integrate my custom loss functions. 2 great features were the web client that collect data from human demonstrations and the data logger which tracks and graphs results in real-time, both of which were super easy to run.”

Congratulations again to both teams!

We would like to thank the organizers of the ALA Workshop for their support of the Challenge. We are happy to see the wide range of interesting ALA projects featured this year. We encourage you to check out Cogment.ai to see if the platform can help your future research. If you think it can, please feel free to connect with us anytime.