Negotiation Game deterministic state

[Linwenye]

Hi! As for the Negotiation game, I was expecting for the randomness of the initial game state, however, after conducting the code

    game = pyspiel.load_game("negotiation")
    state = game.new_initial_state()
    if state.is_chance_node():
        # Sample a chance event outcome.
        outcomes_with_probs = state.chance_outcomes()
        action_list, prob_list = zip(*outcomes_with_probs)
        action = np.random.choice(action_list, p=prob_list)
        state.apply_action(action)
        print(state)

It always return the same state. Why this happen?

[lanctot]

Ah yes, that's because the game has chance mode kExplicitStochastic, which is explained in a comment here:

open_spiel/open_spiel/spiel.h

Line 81 in 77a03df

kSampledStochastic, // At least one chance node with non-deterministic

Essentially it means that there is a single chance outcome and the randomness is applied when you apply the action (so the RNG is stored internally in the game object).

If you don't pass in the rng_seed parameter, it defaults to -1 (see negotiation.h) and then uses the default one for the Mersenne twister, see here:

open_spiel/open_spiel/games/negotiation/negotiation.cc

Line 512 in 77a03df

rng_(new std::mt19937(seed_ >= 0 ? seed_ : std::mt19937::default_seed)) {

(note: the documentation in the header should be updated, I'll fix that).

So it results in the same sequence of samples (after loading the game). Note that you don't need to load the game every time you want to create a new state (and you should only load the game once).

One fix is to set the seed yourself: e.g. pyspiel.load_game("negotiation(rng_seed=3279011)"). If you want the usual explict stochastic chance mode, there is another similar game called "bargaining" which does it using explicit stochastic.