def test_portfolio_weights_get_updated_by_predictions_up_to_one_after_the_batch(memory):
record(memory, 4)
b = get_stable_batch(memory, 2, 1)
b.predictions = [[0.0, 0.0]] * 2
memory.update(b)
b = get_stable_batch(memory, 2, 7)
assert_weights([[0.0]] * 2, b.weights)
def test_portfolio_weights_of_a_batch_can_be_updated_with_predictions(memory):
seed = int(time.time()) # should be stable in a random environment so select a random seed
record(memory, 100)
b = get_stable_batch(memory, 3, seed)
old_w = list(b.weights)
b.predictions = [[0.0, 0.0]] * 3
memory.update(b)
b = get_stable_batch(memory, 3, seed)
assert_weights([old_w[0]] + [[0.0]] * 2, b.weights)
def test_fpm_memory_restores_with_correctly_when_exceeding_capacity(
save_file, default):
cfg = default.window(1).size(2).cfg
m_saved = make_memory(cfg)
record(m_saved, 3)
m_saved.save(save_file)
assert_batch(batch(2.0, 3.0), get_stable_batch(m_saved, 1, 11))
m_restored = make_memory(cfg)
m_restored.restore(save_file)
assert_batch(batch(2.0, 3.0), get_stable_batch(m_restored, 1, 11))
def test_batch_selection_follows_a_geometrically_decaying_distribution(memory):
np.random.seed(7)
memory.beta = 0.5
records = 6
record(memory, records)
distribution = [0] * (records - 2)
n = 1000
for _ in range(0, n):
distribution[identify_state(memory.get_random_batch(2)[0]) - 1] += 1
distribution[:] = [p / n for p in distribution]
assert pytest.approx([0.125, 0.125, 0.25, 0.5], 0.1) == distribution
def test_drop_history_when_memory_capacity_is_reached(memory):
memory.capacity = 2
record(memory, 3)
assert_batch(batch(2.0, 3.0), get_stable_batch(memory, 1, 11))
def test_portfolio_weight_update_is_clamped_to_record_size(memory):
record(memory, 2)
b = get_stable_batch(memory, 2, 1)
b.predictions = [[0.0, 0.0]] * 2
memory.update(b)