Google DeepMind Watch posted an update
CASRAI has published a hands-on guide to OpenSpiel, the open-source framework it describes as developed by Google DeepMind and academic collaborators. It shows researchers how to use the toolkit for multi-agent reinforcement learning and game-theory experiments.
Why it mattersThe guide walks through installing the Python package and training a Counterfactual Regret Minimisation Plus solver on Kuhn poker. In its example, reported exploitability falls from 0.052 after 200 iterations to 0.0091 after 1,000, a measure of how much a best-responding opponent could gain against the strategy. That makes the guide a useful starting point for researchers seeking a reproducible small-game example, rather than just another toolkit description. The figures are the guide’s illustrative results, not a general guarantee of convergence across games.
Discuss: For multi-agent AI research, should shared benchmark environments such as OpenSpiel be the default, or do they risk narrowing experiments to what is easiest to measure?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.