Marc Lanctot

Hi @dpmaloney, yeah what I'd really love is to have a data structure to represent sequence-form strategies (much like we have for behavioral strategies, e.g. `TabularPolicy`). But even separately from...

I think that documentation is referring to the Tensorflow-based C++ and Python versions, we should update it. There's an example evaluation against MCTS here: https://github.com/deepmind/open_spiel/blob/f522d174a1e2e8fbaf2007294985869d6e520669/open_spiel/algorithms/alpha_zero_torch/alpha_zero.cc#L248, but we should also have...

No that is what I meant.. did not realize it already exists!

Thanks @BluemlJ. I agree Stratego is a great game for RL research. However, we cannot import it into our repos without explicit permission from the publishers because the game is...

Ok, great. Yes permissions have been obtained for hosting implementations of games before, e.g. for the Hanabi Learning Environment. It does take some time and the chance of getting them...

Hmm, we don't normally serialize the solver -- does the above work for a different game, e.g. leduc_poker?

Can you try serializing the game or the state separately from the solver.. does that work?

I have a suspicion. The serialization requires reconstructing the state via the game string, and you're loading it by specifying a string that contains newlines, which works fine based on...

I think this might be easily fixed by using a custom serializer for universal poker, which I believe has been on the TODO list for quite some time. I can...

I took a quick look. This is non-trivial to fix properly because it's currently impossible to have custom *game* serializers/deserializers (having custom *state* serializers / deserializers is easy). We should...