"""Tests for reach.py. Run: python -m pytest -q""" import numpy as np import pytest import reach from reach import MLP, behaviour_cloning, clip_step, dagger, distribution_shift, evaluate, expert, rollout def test_clip_step_limits_length() -> None: a = clip_step(np.array([[3.0, 4.0], [0.01, 0.0]])) assert np.allclose(np.linalg.norm(a, axis=1), [reach.MAX_STEP, 0.01]) def test_expert_starts_upwards_and_pulls_back_to_arc() -> None: assert np.allclose(expert(np.array([-1.0, 0.0])), [0.0, reach.SPEED]) inside = expert(np.array([0.0, 0.8])) # below the top of the arc assert inside[1] > 0 # pushed outwards outside = expert(np.array([0.0, 1.2])) assert outside[1] < 0 # pulled inwards def test_expert_reaches_goal() -> None: r = evaluate(expert) assert r["success"] == 1.0 assert r["arc_error"] < 0.03 def test_rollout_is_reproducible() -> None: a = rollout(expert, np.random.default_rng(5)) b = rollout(expert, np.random.default_rng(5)) assert a.shape == (reach.HORIZON + 1, 2) assert np.array_equal(a, b) def test_mlp_fits_a_smooth_function() -> None: rng = np.random.default_rng(0) X = rng.uniform(-1, 1, size=(200, 2)) Y = np.column_stack([np.sin(X[:, 0]), X[:, 0] * X[:, 1]]) loss = MLP(seed=0).fit(X, Y, epochs=1500) assert loss < 1e-3 @pytest.fixture(scope="module") def trained() -> dict[str, object]: bc = behaviour_cloning(35, seed=0) dag, history = dagger(5, 6, 5, seed=0) return {"bc": bc, "dag": dag, "history": history} def test_dagger_beats_bc_with_the_same_number_of_labelled_episodes(trained: dict) -> None: r_bc = evaluate(trained["bc"]) r_dag = evaluate(trained["dag"]) assert r_dag["arc_error"] < 0.6 * r_bc["arc_error"] assert r_dag["final_dist"] < r_bc["final_dist"] assert r_dag["success"] >= r_bc["success"] def test_dagger_improves_on_its_first_round(trained: dict) -> None: history = trained["history"] assert len(history) == 7 assert history[-1] < 0.5 * history[0] # round 0 is behaviour cloning on 5 demos def test_bc_suffers_distribution_shift(trained: dict) -> None: on_demo, on_own = distribution_shift(trained["bc"]) assert on_own > 3 * on_demo on_demo_d, on_own_d = distribution_shift(trained["dag"]) assert on_own_d < 2 * on_demo_d