Skip to content

PyDiffGame is a Python implementation of a Nash Equilibrium solution to Differential Games, based on a reduction of Game Hamilton-Bellman-Jacobi (GHJB) equations to Game Algebraic and Differential Riccati equations, associated with Multi-Objective Dynamical Control Systems

License

Notifications You must be signed in to change notification settings

krichelj/PyDiffGame

Repository files navigation

Logo

pages-build-deployment License: MIT

What is this?

PyDiffGame is a Python implementation of a Nash Equilibrium solution to Differential Games, based on a reduction of Game Hamilton-Bellman-Jacobi (GHJB) equations to Game Algebraic and Differential Riccati equations, associated with Multi-Objective Dynamical Control Systems
The method relies on the formulation given in:

  • The thesis work "Differential Games for Compositional Handling of Competing Control Tasks" (Research Gate)

  • The conference article "Composition of Dynamic Control Objectives Based on Differential Games" (IEEE | Research Gate)

If you use this work, please cite our paper:

@conference{med_paper,  
  author={Kricheli, Joshua Shay and Sadon, Aviran and Arogeti, Shai and Regev, Shimon and Weiss, Gera},
  booktitle={29th Mediterranean Conference on Control and Automation (MED)}, 
  title={{Composition of Dynamic Control Objectives Based on Differential Games}}, 
  year={2021},
  pages={298-304},
  doi={10.1109/MED51440.2021.9480269}}

Installation

To install this package run this from the command prompt:

pip install PyDiffGame

The package was tested for Python >= 3.10, along with the listed packages versions in requirments.txt

Input Parameters

The package defines an abstract class PyDiffGame.py. An object of this class represents an instance of differential game. The input parameters to instantiate a PyDiffGame object are:

  • A : np.array of shape $(n,n)$

System dynamics matrix

  • B : np.array of shape $(n, m_1 + ... + m_N)$, optional

Input matrix for all virtual control objectives

  • Bs : Sequence of np.array objects of len $(N)$, each array $B_i$ of shape $(n,m_i)$, optional

Input matrices for each virtual control objective

  • Qs : Sequence of np.array objects of len $(N)$, each array $Q_i$ of shape $(n,n)$, optional

State weight matrices for each virtual control objective

  • Rs : Sequence of np.array objects of len $(N)$, each array $R_i$ of shape $(m_i,m_i)$, optional

Input weight matrices for each virtual control objective

  • Ms : Sequence of np.array objects of len $(N)$, each array $M_i$ of shape $(m_i,m)$, optional

Decomposition matrices for each virtual control objective

  • objectives : Sequence of Objective objects of len $(N)$, each $O_i$ specifying $Q_i, R_i$ and $M_i$, optional

Desired objectives for the game

  • x_0 : np.array of len $(n)$, optional

Initial state vector

  • x_T : np.array of len $(n)$, optional

Final state vector, in case of signal tracking

  • T_f : positive float, optional

System dynamics horizon. Should be given in the case of finite horizon

  • P_f : list of np.array objects of len $(N)$, each array $P_{f_i}$ of shape $(n,n)$, optional, default = uncoupled solution of scipy's solve_are

Final condition for the Riccati equation array. Should be given in the case of finite horizon

  • state_variables_names : Sequence of str objects of len $(N)$, optional

The state variables' names to display when plotting

  • show_legend : boolean, optional

Indicates whether to display a legend in the plots

  • state_variables_names : Sequence of str objects of len $(n)$, optional

The state variables' names to display

  • epsilon_x : float in the interval $(0,1)$, optional

Numerical convergence threshold for the state vector of the system

  • epsilon_P : float in the interval $(0,1)$, optional

Numerical convergence threshold for the matrices P_i

  • L : positive int, optional

Number of data points

  • eta : positive int, optional

The number of last matrix norms to consider for convergence

  • debug : boolean, optional

Indicates whether to display debug information

Tutorial

To demonstrate the use of the package, we provide a few running examples. Consider the following system of masses and springs:

The performance of the system under the use of the suggested method is compared with that of a Linear Quadratic Regulator (LQR). For that purpose, class named PyDiffGameLQRComparison is defined. A comparison of a system should subclass this class. As an example, for the masses and springs system, consider the following instantiation of an MassesWithSpringsComparison object:

import numpy as np
from PyDiffGame.examples.MassesWithSpringsComparison import MassesWithSpringsComparison

N = 2
k = 10
m = 50
r = 1
epsilon_x = 10e-8
epsilon_P = 10e-8
q = [[500, 2000], [500, 250]]

x_0 = np.array([10 * i for i in range(1, N + 1)] + [0] * N)
x_T = x_0 * 10 if N == 2 else np.array([(10 * i) ** 3 for i in range(1, N + 1)] + [0] * N)
T_f = 25

masses_with_springs = MassesWithSpringsComparison(N=N,
                                                  m=m,
                                                  k=k,
                                                  q=q,
                                                  r=r,
                                                  x_0=x_0,
                                                  x_T=x_T,
                                                  T_f=T_f,
                                                  epsilon_x=epsilon_x,
                                                  epsilon_P=epsilon_P)

Consider the constructor of the class MassesWithSpringsComparison:

import numpy as np
from typing import Sequence, Optional

from PyDiffGame.PyDiffGame import PyDiffGame
from PyDiffGame.PyDiffGameLQRComparison import PyDiffGameLQRComparison
from PyDiffGame.Objective import GameObjective, LQRObjective


class MassesWithSpringsComparison(PyDiffGameLQRComparison):
    def __init__(self,
                 N: int,
                 m: float,
                 k: float,
                 q: float | Sequence[float] | Sequence[Sequence[float]],
                 r: float,
                 Ms: Optional[Sequence[np.array]] = None,
                 x_0: Optional[np.array] = None,
                 x_T: Optional[np.array] = None,
                 T_f: Optional[float] = None,
                 epsilon_x: Optional[float] = PyDiffGame.epsilon_x_default,
                 epsilon_P: Optional[float] = PyDiffGame.epsilon_P_default,
                 L: Optional[int] = PyDiffGame.L_default,
                 eta: Optional[int] = PyDiffGame.eta_default):
        I_N = np.eye(N)
        Z_N = np.zeros((N, N))

        M_masses = m * I_N
        K = k * (2 * I_N - np.array([[int(abs(i - j) == 1) for j in range(N)] for i in range(N)]))
        M_masses_inv = np.linalg.inv(M_masses)

        M_inv_K = M_masses_inv @ K

        if Ms is None:
            eigenvectors = np.linalg.eig(M_inv_K)[1]
            Ms = [eigenvector.reshape(1, N) for eigenvector in eigenvectors]

        A = np.block([[Z_N, I_N],
                      [-M_inv_K, Z_N]])
        B = np.block([[Z_N],
                      [M_masses_inv]])

        Qs = [np.diag([0.0] * i + [q] + [0.0] * (N - 1) + [q] + [0.0] * (N - i - 1))
              if isinstance(q, (int, float)) else
              np.diag([0.0] * i + [q[i]] + [0.0] * (N - 1) + [q[i]] + [0.0] * (N - i - 1)) for i in range(N)]

        M = np.concatenate(Ms,
                           axis=0)

        assert np.all(np.abs(np.linalg.inv(M) - M.T) < 10e-12)

        Q_mat = np.kron(a=np.eye(2),
                        b=M)

        Qs = [Q_mat.T @ Q @ Q_mat for Q in Qs]

        Rs = [np.array([r])] * N
        R_lqr = 1 / 4 * r * I_N
        Q_lqr = q * np.eye(2 * N) if isinstance(q, (int, float)) else np.diag(2 * q)

        state_variables_names = ['x_{' + str(i) + '}' for i in range(1, N + 1)] + \
                                ['\\dot{x}_{' + str(i) + '}' for i in range(1, N + 1)]
        args = {'A': A,
                'B': B,
                'x_0': x_0,
                'x_T': x_T,
                'T_f': T_f,
                'state_variables_names': state_variables_names,
                'epsilon_x': epsilon_x,
                'epsilon_P': epsilon_P,
                'L': L,
                'eta': eta}

        lqr_objective = [LQRObjective(Q=Q_lqr,
                                      R_ii=R_lqr)]
        game_objectives = [GameObjective(Q=Q,
                                         R_ii=R,
                                         M_i=M_i) for Q, R, M_i in zip(Qs, Rs, Ms)]
        games_objectives = [lqr_objective,
                            game_objectives]

        super().__init__(args=args,
                         M=M,
                         games_objectives=games_objectives,
                         continuous=True)

Finally, consider calling the masses_with_springs object as follows:

output_variables_names = ['$\\frac{x_1 + x_2}{\\sqrt{2}}$',
                          '$\\frac{x_2 - x_1}{\\sqrt{2}}$',
                          '$\\frac{\\dot{x}_1 + \\dot{x}_2}{\\sqrt{2}}$',
                          '$\\frac{\\dot{x}_2 - \\dot{x}_1}{\\sqrt{2}}$']

masses_with_springs(plot_state_spaces=True,
                    plot_Mx=True,
                    output_variables_names=output_variables_names,
                    save_figure=True)

Refer This will result in the following plot that compares the two systems performance for a differential game vs an LQR:

And when tweaking the weights by setting

qs = [[500, 5000]]

we have:

Acknowledgments

This research was supported in part by the Leona M. and Harry B. Helmsley Charitable Trust through the Agricultural, Biological and Cognitive Robotics Initiative and by the Marcus Endowment Fund both at Ben-Gurion University of the Negev, Israel. This research was also supported by The Israeli Smart Transportation Research Center (ISTRC) by The Technion and Bar-Ilan Universities, Israel.

    Helmsley Charitable Trust