edsger 0.1.1__pp39-pypy39_pp73-macosx_10_9_x86_64.whl

edsger/spiess_florian.pyx

@@ -0,0 +1,368 @@
+"""
+An implementation of Spiess and Florian's hyperpath generating algorithm.
+
+reference: Spiess, H. and Florian, M. (1989). Optimal strategies: A new
+assignment model for transit networks. Transportation Research Part B 23(2),
+83-102.
+
+cpdef function:
+
+- compute_SF_in
+    Compute Spiess & Florian optimal strategy (all-to-one).
+
+cdef functions:
+
+- _SF_in_first_pass_full
+- _SF_in_second_pass
+
+"""
+
+import numpy as np
+cimport numpy as cnp
+
+from edsger.commons import DTYPE_PY, DTYPE_INF_PY
+from edsger.commons cimport (
+    DTYPE_INF, UNLABELED, SCANNED, DTYPE_t, ElementState)
+cimport edsger.pq_4ary_dec_0b as pq  # priority queue
+
+
+cpdef void compute_SF_in(
+    cnp.uint32_t[::1] csc_indptr,
+    cnp.uint32_t[::1] csc_edge_idx,
+    DTYPE_t[::1] c_a_vec,
+    DTYPE_t[::1] f_a_vec,
+    cnp.uint32_t[::1] tail_indices,
+    cnp.uint32_t[::1] head_indices,
+    cnp.uint32_t[::1] demand_indices,
+    DTYPE_t[::1] demand_values,
+    DTYPE_t[::1] v_a_vec,
+    DTYPE_t[::1] u_i_vec,
+    int vertex_count,
+    int dest_vert_index,
+):
+    """Compute Spiess & Florian optimal strategy (all-to-one).
+    """
+
+    cdef:
+        int edge_count = tail_indices.shape[0]
+
+    # initialization
+    u_i_vec[<size_t>dest_vert_index] = 0.0
+
+    # vertex properties
+    # vertex frequency (inverse of the maximum delay)
+    f_i_vec = np.zeros(vertex_count, dtype=DTYPE_PY)
+    u_j_c_a_vec = DTYPE_INF_PY * np.ones(edge_count, dtype=DTYPE_PY)
+    v_i_vec = np.zeros(vertex_count, dtype=DTYPE_PY)  # vertex volume
+
+    # edge properties
+    h_a_vec = np.zeros(edge_count, dtype=bool)  # edge belonging to hyperpath
+
+    # first pass #
+    # ---------- #
+
+    _SF_in_first_pass_full(
+        csc_indptr,
+        csc_edge_idx,
+        c_a_vec,
+        f_a_vec,
+        tail_indices,
+        u_i_vec,
+        f_i_vec,
+        u_j_c_a_vec,
+        h_a_vec,
+        dest_vert_index
+    )
+
+    # second pass #
+    # ----------- #
+
+    cdef:
+        DTYPE_t u_r, _v_a_new, _v_i, u_i
+        size_t i, h_a_count
+        cnp.uint32_t vert_idx
+
+    v_i_vec = np.zeros(vertex_count, dtype=DTYPE_PY)  # vertex volume
+
+    u_r = DTYPE_INF_PY
+    for i, vert_idx in enumerate(demand_indices):
+
+        v_i_vec[<size_t>vert_idx] = demand_values[i]
+        u_i = u_i_vec[<size_t>vert_idx]
+
+        if u_i < u_r:
+
+            u_r = u_i
+
+    # if the destination can be reached from any of the origins
+    if u_r < DTYPE_INF_PY:
+
+        # make sure f_i values are not zero
+        f_i_vec = np.where(
+            f_i_vec < MIN_FREQ_PY, MIN_FREQ_PY, f_i_vec
+        )
+
+        # sort the links with descreasing order of u_j + c_a
+        h_a_count = h_a_vec.sum()
+        masked_a = np.ma.array(-u_j_c_a_vec, mask=~h_a_vec)
+        edge_indices = np.argsort(masked_a).astype(np.uint32)
+
+        _SF_in_second_pass(
+            edge_indices,
+            tail_indices,
+            head_indices,
+            v_i_vec,
+            v_a_vec,
+            f_i_vec,
+            f_a_vec,
+            h_a_count
+        )
+
+
+cdef void _SF_in_first_pass_full(
+    cnp.uint32_t[::1] csc_indptr,
+    cnp.uint32_t[::1] csc_edge_idx,
+    DTYPE_t[::1] c_a_vec,
+    DTYPE_t[::1] f_a_vec,
+    cnp.uint32_t[::1] tail_indices,
+    DTYPE_t[::1] u_i_vec,
+    DTYPE_t[::1] f_i_vec,
+    DTYPE_t[::1] u_j_c_a_vec,
+    cnp.uint8_t[::1] h_a_vec,
+    int dest_vert_index,
+) nogil:
+    """SF in first pass.
+
+    Note : all vertices are visited.
+    """
+
+    cdef:
+        int edge_count = tail_indices.shape[0]
+        pq.PriorityQueue pqueue
+        ElementState edge_state
+        size_t i, edge_idx, tail_vert_idx
+        DTYPE_t u_j_c_a, u_i, f_i, beta, u_i_new, f_a
+
+    # initialization of the heap elements
+    # all nodes have INFINITY key and UNLABELED state
+    pq.init_pqueue(&pqueue, <size_t>edge_count, <size_t>edge_count)
+
+    # only the incoming edges of the target vertex are inserted into the
+    # priority queue
+    for i in range(<size_t>csc_indptr[<size_t>dest_vert_index],
+                   <size_t>csc_indptr[<size_t>(dest_vert_index + 1)]):
+        edge_idx = csc_edge_idx[i]
+        pq.insert(&pqueue, edge_idx, c_a_vec[edge_idx])
+        u_j_c_a_vec[edge_idx] = c_a_vec[edge_idx]
+
+    # first pass
+    while pqueue.size > 0:
+
+        edge_idx = pq.extract_min(&pqueue)
+        u_j_c_a = pqueue.Elements[edge_idx].key
+        tail_vert_idx = <size_t>tail_indices[edge_idx]
+        u_i = u_i_vec[tail_vert_idx]
+
+        if u_i >= u_j_c_a:
+
+            f_i = f_i_vec[tail_vert_idx]
+
+            # compute the beta coefficient
+            if (u_i < DTYPE_INF) | (f_i > 0.0):
+
+                beta = f_i * u_i
+
+            else:
+
+                beta = 1.0
+
+            # update u_i
+            f_a = f_a_vec[edge_idx]
+            u_i_new = (beta + f_a * u_j_c_a) / (f_i + f_a)
+            u_i_vec[tail_vert_idx] = u_i_new
+
+            # update f_i
+            f_i_vec[tail_vert_idx] = f_i + f_a
+
+            # add the edge to hyperpath
+            h_a_vec[edge_idx] = 1
+
+        else:
+
+            u_i_new = u_i
+
+        # loop on incoming edges
+        for i in range(<size_t>csc_indptr[tail_vert_idx],
+                       <size_t>csc_indptr[tail_vert_idx + 1]):
+
+            edge_idx = csc_edge_idx[i]
+            edge_state = pqueue.Elements[edge_idx].state
+
+            if edge_state != SCANNED:
+
+                # u_j of current edge = u_i of outgoing edge
+                u_j_c_a = u_i_new + c_a_vec[edge_idx]
+
+                if edge_state == UNLABELED:
+
+                    pq.insert(&pqueue, edge_idx, u_j_c_a)
+                    u_j_c_a_vec[edge_idx] = u_j_c_a
+
+                elif (pqueue.Elements[edge_idx].key > u_j_c_a):
+
+                    pq.decrease_key(&pqueue, edge_idx, u_j_c_a)
+                    u_j_c_a_vec[edge_idx] = u_j_c_a
+
+    pq.free_pqueue(&pqueue)
+
+
+cdef void _SF_in_second_pass(
+    cnp.uint32_t[::1] edge_indices,
+    cnp.uint32_t[::1] tail_indices,
+    cnp.uint32_t[::1] head_indices,
+    DTYPE_t[::1] v_i_vec,
+    DTYPE_t[::1] v_a_vec,
+    DTYPE_t[::1] f_i_vec,
+    DTYPE_t[::1] f_a_vec,
+    size_t h_a_count
+) nogil:
+    """SF in second_pass.
+    """
+
+    cdef:
+        size_t i, edge_idx, vert_idx
+        DTYPE_t v_i, f_i, f_a, v_a_new
+
+    for i in range(h_a_count):
+
+        edge_idx = <size_t>edge_indices[i]
+        vert_idx = <size_t>tail_indices[edge_idx]
+
+        v_i = v_i_vec[vert_idx]
+        f_i = f_i_vec[vert_idx]
+        f_a = f_a_vec[edge_idx]
+
+        # update v_a
+        v_a_new = v_i * f_a / f_i
+        v_a_vec[edge_idx] = v_a_new
+        v_i_vec[<size_t>head_indices[edge_idx]] += v_a_new
+
+
+# ============================================================================ #
+# tests                                                                        #
+# ============================================================================ #
+
+
+from edsger.commons import MIN_FREQ_PY, INF_FREQ_PY
+
+cpdef compute_SF_in_01():
+    """
+    Single edge network.
+
+    This network has 1 edge and 2 vertices.
+    """
+
+    volume = 1.0
+
+    f_a = MIN_FREQ_PY
+
+    csc_indptr = np.array([0, 0, 1], dtype=np.uint32)
+    csc_edge_idx = np.array([0], dtype=np.uint32)
+    c_a_vec = np.array([1.0], dtype=DTYPE_PY)
+    f_a_vec = np.array([f_a], dtype=DTYPE_PY)
+    v_a_vec = np.array([0.0], dtype=DTYPE_PY)
+    tail_indices = np.array([0], dtype=np.uint32)
+    head_indices = np.array([1], dtype=np.uint32)
+    demand_indices = np.array([0], dtype=np.uint32)
+    demand_values = np.array([volume], dtype=DTYPE_PY)
+    vertex_count = 2
+    u_i_vec = DTYPE_INF_PY * np.ones(vertex_count, dtype=DTYPE_PY)
+    dest_vert_index = 1
+
+    compute_SF_in(
+        csc_indptr,
+        csc_edge_idx,
+        c_a_vec,
+        f_a_vec,
+        tail_indices,
+        head_indices,
+        demand_indices,
+        demand_values,
+        v_a_vec,
+        u_i_vec,
+        vertex_count,
+        dest_vert_index,
+    )
+
+    assert v_a_vec[0] == volume
+    assert v_a_vec.shape[0] == 1
+
+    f_a = INF_FREQ_PY
+    f_a_vec = np.array([f_a], dtype=DTYPE_PY)
+    u_i_vec = DTYPE_INF_PY * np.ones(vertex_count, dtype=DTYPE_PY)
+
+    compute_SF_in(
+        csc_indptr,
+        csc_edge_idx,
+        c_a_vec,
+        f_a_vec,
+        tail_indices,
+        head_indices,
+        demand_indices,
+        demand_values,
+        v_a_vec,
+        u_i_vec,
+        vertex_count,
+        dest_vert_index,
+    )
+
+    assert v_a_vec[0] == volume
+    assert v_a_vec.shape[0] == 1
+
+
+cpdef compute_SF_in_02():
+    """
+    Two edges network.
+
+    This network has 2 edges and 2 vertices.
+    """
+
+    volume = 1.0
+
+    csc_indptr = np.array([0, 0, 2], dtype=np.uint32)
+    csc_edge_idx = np.array([0, 1], dtype=np.uint32)
+    c_a_vec = np.array([1.0, 1.0], dtype=DTYPE_PY)
+    f_a_vec = np.array([2.0, 1.0], dtype=DTYPE_PY)
+    v_a_vec = np.array([0.0, 0.0], dtype=DTYPE_PY)
+    tail_indices = np.array([0, 0], dtype=np.uint32)
+    head_indices = np.array([1, 1], dtype=np.uint32)
+    demand_indices = np.array([0], dtype=np.uint32)
+    demand_values = np.array([volume], dtype=DTYPE_PY)
+    vertex_count = 2
+    u_i_vec = DTYPE_INF_PY * np.ones(vertex_count, dtype=DTYPE_PY)
+    dest_vert_index = 1
+
+    compute_SF_in(
+        csc_indptr,
+        csc_edge_idx,
+        c_a_vec,
+        f_a_vec,
+        tail_indices,
+        head_indices,
+        demand_indices,
+        demand_values,
+        v_a_vec,
+        u_i_vec,
+        vertex_count,
+        dest_vert_index,
+    )
+
+    assert v_a_vec[0] == 2. / 3.
+    assert v_a_vec[1] == 1. / 3.
+    assert v_a_vec.shape[0] == 2
+
+
+# author : Francois Pacull
+# copyright : Architecture & Performance
+# email: francois.pacull@architecture-performance.fr
+# license : MIT