fictitious-play.rst

Fictitious play

The fictitious play algorithm implemented in Nashpy is based on the one described in [Fudenberg1998].

The algorithm is as follows:

For a game (A, B) ∈ ℝ^m × n define κ_tⁱ : S^− 1 → ℕ to be a function that in a given time interval t for a player i maps a strategy s from the opponent's strategy space S^− 1 to a number of total times the opponent has played s.

Thus:

$$\begin{aligned} \kappa_t^{i}(s^{-i}) = \kappa_{t-1}(s^{-i}) + \begin{cases} 1,& \text{ if }s^{-i}_{t-1}=s^{-i}\\\ 0,& \text{ otherwise} \end{cases} \end{aligned}$$

In practice:

κ_t¹ ∈ ℤⁿ  κ_t² ∈ ℤ^m

At stage t, each player assumes their opponent is playing a mixed strategy based on κ_t − 1:

$$\frac{\kappa_{t-1}}{\sum\kappa_{t-1}}$$

They calculate the expected value of each strategy, which is equivalent to:

s_t¹ ∈ argmax_s ∈ S1Aκ_t − 1²  s_t² ∈ argmax_s ∈ S2B^Tκ_t − 1¹

In the case of multiple best responses, a random choice is made.

Discussion

Note that this algorithm will not always converge and sometimes it depends on the form of the game.

For example:

>>> import numpy as np
>>> import nashpy as nash
>>> A = np.array([[0, 1, 0], [0, 0, 1], [1, 0, 0]])
>>> B = np.array([[0, 0, 1], [1, 0, 0], [0, 1, 0]])
>>> game = nash.Game(A, B)
>>> iterations = 10000
>>> np.random.seed(0)
>>> play_counts = tuple(game.fictitious_play(iterations=iterations))
>>> play_counts[-1]
[array([5464., 1436., 3100.]), array([2111., 4550., 3339.])]

We can visualise the lack of convergence:

>>> import matplotlib.pyplot as plt
>>> plt.figure() # doctest: +SKIP
>>> probabilities = [row_play_counts / np.sum(row_play_counts) for row_play_counts, col_play_counts in play_counts]
>>> for number, strategy in enumerate(zip(*probabilities)):
...     plt.plot(strategy, label=f"$s_{number}$")  # doctest: +SKIP
>>> plt.xlabel("Iteration")  # doctest: +SKIP
>>> plt.ylabel("Probability")  # doctest: +SKIP
>>> plt.title("Actions taken by row player")  # doctest: +SKIP
>>> plt.legend()  # doctest: +SKIP

If we modify the game slightly we obtain a different outcome:

>>> A = np.array([[1 / 2, 1, 0], [0, 1 / 2, 1], [1, 0, 1 / 2]])
>>> B = np.array([[1 / 2, 0, 1], [1, 1 / 2, 0], [0, 1, 1 / 2]])
>>> game = nash.Game(A, B)
>>> np.random.seed(0)
>>> play_counts = tuple(game.fictitious_play(iterations=iterations))
>>> play_counts[-1]
[array([3290., 3320., 3390.]), array([3356., 3361., 3283.])]

With a clear convergence now visible:

>>> import matplotlib.pyplot as plt
>>> plt.figure() # doctest: +SKIP
>>> probabilities = [row_play_counts / np.sum(row_play_counts) for row_play_counts, col_play_counts in play_counts]
>>> for number, strategy in enumerate(zip(*probabilities)):
...     plt.plot(strategy, label=f"$s_{number}$")  # doctest: +SKIP
>>> plt.xlabel("Iteration")  # doctest: +SKIP
>>> plt.ylabel("Probability")  # doctest: +SKIP
>>> plt.title("Actions taken by row player")  # doctest: +SKIP
>>> plt.legend()  # doctest: +SKIP