Add PIR (Physics Intermediate Representation) symbolic regression method

algorithms/pir/install.sh

@@ -0,0 +1,2 @@
+#!/bin/bash
+pip install "git+https://github.com/Qazi-pk/[email protected]"

algorithms/pir/metadata.yml

@@ -0,0 +1,19 @@
+# PIR — Physics Intermediate Representation
+name: PIR
+authors:
+  - name: Qazi Hanif
+    email: [email protected]
+email: [email protected]
+url: https://github.com/Qazi-pk/physics-engine
+key: PIR
+paper:
+  title: "PIR: Physics Intermediate Representation for Automated Discovery of Physical Laws"
+  url: https://doi.org/10.5281/zenodo.21351039
+description: >
+  Classical symbolic regression via monomial-basis search with log-linearization
+  gate for power-law detection (F3 gate), pairwise structure decomposition,
+  RANSAC, sparse regression, and iterative residual refinement with Occam
+  complexity penalty. No neural components. Verified blind Feynman Tier A:
+  12/44 EXACT (zero seed wobble, v3.4). Secondary: +12/44 FORM_NUMERIC
+  (correct functional form, transcendental constant as decimal, reported
+  separately and never summed into the primary figure).

algorithms/pir/requirements.txt

@@ -0,0 +1 @@
+POT

experiment/methods/pir/__init__.py

@@ -0,0 +1 @@
+# PIR method package

experiment/methods/pir/regressor.py

@@ -0,0 +1,181 @@
+"""
+SRBench method: PIR (Physics Intermediate Representation) — classical engine.
+
+Importable module for the SRBench harness (imported as `methods.pir`).
+Wraps the public, pip-installable classical PIR engine (installed via
+algorithms/pir/install.sh) in the interface SRBench expects:
+
+    est                    a scikit-learn-compatible Regressor instance
+    model(est, X=None)     returns a sympy-parseable string for the fitted model
+    complexity(est)        returns an integer complexity count for the model
+    eval_kwargs            method-specific args forwarded to evaluate_model.py
+
+No neural components. Vanilla classical PIR only — this is the configuration
+that produced the verified blind Tier A baseline:
+
+    12/44 EXACT  (zero seed wobble, v3.4)
+    +12/44 FORM_NUMERIC secondary (correct functional form, transcendental
+                constant folded as decimal — reported separately, never summed)
+
+Engine: https://github.com/Qazi-pk/physics-engine (MIT, tag v3.4.1)
+Paper: https://doi.org/10.5281/zenodo.21351039
+"""
+
+import re
+
+import numpy as np
+
+from sklearn.base import BaseEstimator, RegressorMixin
+
+# The classical engine. algorithms/pir/install.sh pip-installs this from the
+# public repo at tag v3.4.1.
+from physics_engine.sklearn_adapter import PIRRegressor
+
+
+# --- Vanilla config that produced the verified blind Tier A result ------------
+_PIR_VANILLA_CONFIG = dict(
+    enforce_dimensions=False,        # blind sweep ran with dim-filter OFF
+    allowed_powers=[1, 2],           # powers 1,2 only (structural cap)
+    include_pairwise_products=True,  # pairwise on; no 3-var assembly
+    use_ransac=True,
+    use_residual=True,
+    use_sparse=True,
+    use_ot_loss=False,
+    add_physics_features=False,
+)
+
+
+class PIRClassicRegressor(BaseEstimator, RegressorMixin):
+    """Thin sklearn wrapper around the classical PIRRegressor.
+
+    Adds a `random_state` attribute and a guaranteed-valid fallback model so
+    that model() never raises even if the inner fit fails. Timeout
+    enforcement is left entirely to SRBench's own outer alarm: signal.alarm
+    is process-global, so a second SIGALRM set here would silently overwrite
+    SRBench's own timeout handler (per gAldeia's review on PR #210). The
+    `max_time` constructor kwarg is still accepted, for API compatibility
+    with the harness's test_params override, but is currently unused.
+    """
+
+    def __init__(self, max_time=3600, random_state=None, **pir_kwargs):
+        self.max_time = max_time
+        self.random_state = random_state
+        self.pir_kwargs = {**_PIR_VANILLA_CONFIG, **pir_kwargs}
+
+    def _build(self):
+        kw = dict(self.pir_kwargs)
+        seed = 0 if self.random_state is None else self.random_state
+        try:
+            return PIRRegressor(random_state=seed, **kw)
+        except TypeError:
+            # Engine may not accept random_state; harmless to omit.
+            return PIRRegressor(**kw)
+
+    def fit(self, X, y):
+        # Guarantee a valid model exists before risking a timeout:
+        # constant = mean(y). evaluate_model.py can always score this.
+        y_arr = np.asarray(y, dtype=float).ravel()
+        self._fallback_expr_ = repr(float(np.mean(y_arr))) if y_arr.size else "0.0"
+        self.expr_ = self._fallback_expr_
+        self._inner = self._build()
+
+        # No process-level alarm here — SRBench's own outer timeout
+        # enforces max_time; see class docstring.
+        try:
+            self._inner.fit(X, y)
+            if hasattr(self._inner, "model"):
+                self.expr_ = self._inner.model()
+            elif hasattr(self._inner, "expr_"):
+                self.expr_ = str(self._inner.expr_)
+        except Exception:
+            # Keep the constant fallback already stored in self.expr_.
+            pass
+        self.is_fitted_ = True
+        return self
+
+    def predict(self, X):
+        if hasattr(self, "_inner") and hasattr(self._inner, "predict"):
+            try:
+                return self._inner.predict(X)
+            except Exception:
+                pass
+        # Fallback: constant prediction matching the fallback expression.
+        n = X.shape[0] if hasattr(X, "shape") else len(X)
+        try:
+            val = float(self._fallback_expr_)
+        except (TypeError, ValueError):
+            val = 0.0
+        return np.full(n, val, dtype=float)
+
+    def model(self):
+        return self.expr_
+
+
+# The estimator SRBench will fit.
+est = PIRClassicRegressor(max_time=3600, random_state=None)
+
+
+def model(est, X=None):
+    """Return a sympy-parseable model string with symbols matching X.columns.
+
+    If the engine already emits dataset column names, the string is returned
+    unchanged. Otherwise generic positional names (x_0, x0, X0, ...) are
+    remapped to the dataset columns.
+    """
+    expr = (
+        est.model()
+        if hasattr(est, "model")
+        else str(getattr(est, "expr_", "0.0"))
+    )
+
+    if X is None or not hasattr(X, "columns"):
+        return expr
+
+    cols = list(X.columns)
+    # Word-boundary match, not substring: a column literally named "co"
+    # must not match inside "cos(x_0)" (per gAldeia's review on PR #210).
+    def _col_appears(col_name, s):
+        return re.search(r"\b" + re.escape(str(col_name)) + r"\b", s) is not None
+    # If any real column name already appears, assume names are correct.
+    if any(_col_appears(c, expr) for c in cols):
+        return expr
+
+    # Remap generic positional names -> dataset columns.
+    # reversed() so 'x_1' doesn't clobber the prefix of 'x_10'.
+    for prefix in ("x_", "x", "X_", "X"):
+        if re.search(rf"\b{prefix}\d+\b", expr):
+            mapping = {f"{prefix}{i}": str(k) for i, k in enumerate(cols)}
+            for k, v in reversed(list(mapping.items())):
+                expr = re.sub(rf"\b{re.escape(k)}\b", v, expr)
+            break
+    return expr
+
+
+def complexity(est):
+    """Integer complexity of the fitted model.
+
+    Counts nodes in the expression by splitting on operators and separators,
+    following the convention used by other SRBench methods (e.g. gplearn).
+    SRBench normally uses SymPy's own complexity first; this is only a
+    fallback, but the operator count is included so the fallback is as
+    accurate as possible (per gAldeia's review on PR #210).
+    """
+    expr = model(est)
+    if not expr:
+        return 0
+    # Count operand-like tokens and add back the operators the split removed.
+    tokens = re.split(r"[\s\(\),\+\-\*\/\^]+", expr)
+    operand_count = len([t for t in tokens if t])
+    operator_count = len(re.findall(r"[\+\-\*\/\^]", expr))
+    return operand_count + operator_count
+
+
+# --- forwarded to evaluate_model.py ------------------------------------------
+# CRITICAL: scale_x/scale_y MUST be False. SRBench StandardScales X and y by
+# default, which destroys the units and exact coefficients PIR depends on.
+# `test_params` shortens run-time during CI smoke tests (master-branch signature).
+eval_kwargs = {
+    "scale_x": False,
+    "scale_y": False,
+    "test_params": {"max_time": 60},
+}