theocabannes/open_spiel

OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games.

• 👋 Hi, I’m Theo
• Feel free to find more info on my website

open_spiel's Issues

Issue with player index in Dynamic routing game tests

Cannot compute Nash exploitability with dummy markov dynamic routing game

Exploitability computation raises `RecursionError: maximum recursion depth exceeded`

DynamicRoutingGameTest.test_ext_mccfr_on_turn_based_game_more_complex_network

GameInfo utility_sum should be defined even for non constant-sum-utility game

Ideally, utility_sum should have a default value in case the game is not a constant-sum-utility game.

Simplify exploitability computation when players are symmetric

Provides information tensor value changes when calling convert_to_turn_based

Error in automatic test due to pickle issue

Inconsistent documentation in case no possible actions

In case a player has no possible actions, the documentation is not clear.
ApplyActions states that the player should have the action kInvalidAction: https://github.com/deepmind/open_spiel/blob/53eeb6127578f91e88a6c0451983d10e3509446d/open_spiel/spiel.h#L557.
LegalActions seems to say that the player should have a empty list: https://github.com/deepmind/open_spiel/blob/53eeb6127578f91e88a6c0451983d10e3509446d/open_spiel/spiel.h#L229

However, from generate_playthrough.py, it seems the player should have kInvalidAction as action.

Also, if kInvalidAction is returned then cfr does not work #7.

If empty list is returned then both cfr and NFSP do not work:

Distribution does not sum to one in mean field routing game

It seems that get all states for Mean field games to not see all possible states

To be confirmed, debugged and potentially convert to a upstream issue.

Error in automatic test due to unknown parameter 'players'.

Issue when running NFSP on dummy Markov dynamic routing game

It seems that some legal actions might not be well defined.

Issue with turn-based game conversion from a string.

legal_actions is empty for sequential games created with convert_to_turn_based

Register python_markov_dynamic_routing_game

I succeed in registering markov_dynamic_routing_game by adding it to https://github.com/TheoCabannes/open_spiel/blob/fd2565ad12e125b609fa36abc4d8bcfb00fbbbdf/open_spiel/games/CMakeLists.txt and rebuilding the repo with https://github.com/TheoCabannes/open_spiel/blob/fd2565ad12e125b609fa36abc4d8bcfb00fbbbdf/open_spiel/scripts/build_and_run_tests.sh.

However I did not managed to register python_markov_dynamic_routing_game (which might require changing a python binding file or a make file).

CFR tries to apply actions that are not legal in sequential version of simultaneous game

This is the issue in deepmind#599

CFR does not work currently because it cannot apply invalid actions

In case one player cannot do anything, it is asked to return kInvalidAction (https://github.com/deepmind/open_spiel/blob/53eeb6127578f91e88a6c0451983d10e3509446d/open_spiel/spiel.h#L557).
However, kInvalidAction makes CFR crash (see above and https://github.com/deepmind/open_spiel/blob/53eeb6127578f91e88a6c0451983d10e3509446d/open_spiel/spiel.cc#L317).

MCCFR is not working on current dynamic routing game implementation