Keep track of best seen

ex.py

@@ -20,7 +20,7 @@ prgm = Program(
 outputs = [R8.A]
 test_cases = [TestCase()]
 max_size = 4
-num_iters = 1_000_000  # 10_000_000
+num_iters = 100_000_000
 
 initial_cost = cost(prgm, test_cases, outputs, prgm)
 initial_cycles = prgm.perf()
@@ -44,20 +44,6 @@ end_time = time()
 
 optimized_prgm.display()
 
-optimized_cost = cost(prgm, test_cases, outputs, optimized_prgm)
-optimized_cycles = optimized_prgm.perf()
-
-print(f"Cost: {optimized_cost}")
-print(f"Cycles: {optimized_cycles}")
-
-
-optimized_cost = cost(prgm, test_cases, outputs, optimized_prgm)
-optimized_cycles = optimized_prgm.perf()
-
-print(f"Cost: {optimized_cost}")
-print(f"Cycles: {optimized_cycles}")
-
-
 optimized_cost = cost(prgm, test_cases, outputs, optimized_prgm)
 optimized_cycles = optimized_prgm.perf()
 

gbso/optimize.py

@@ -1,6 +1,6 @@
 from math import exp, log
 from random import random
-from typing import List
+from typing import List, Tuple
 
 from gbso.program.test_case import Output, TestCase, eq_on_testcase
 from gbso.program.mutate import mutate_program
@@ -18,15 +18,16 @@ DEFAULT_PROB_INSN = 0.25
 DEFAULT_PROB_INSN_UNUSED = 0.1
 
 
-def cost(orig_prgm, test_cases, outputs, prgm) -> int:
+def cost(orig_prgm, test_cases, outputs, prgm) -> Tuple[int, bool]:
     c = prgm.perf() - orig_prgm.perf()
+    eq = c == 0
     # print(f"init cost: {c}")
 
     for test_case in test_cases:
         c += eq_on_testcase(orig_prgm, prgm, test_case, outputs)
         # print(f"cost after testcase: {c}")
 
-    return c
+    return c, eq
 
 
 def optimize(
@@ -34,7 +35,7 @@ def optimize(
     max_size: int,
     test_cases: List[TestCase],
     outputs: List[Output],
-    beta: int = 0.75,  # How far away in cost you are allowed to search
+    beta: int = 0.5,  # How far away in cost you are allowed to search
     num_iters: int = DEFAULT_NUM_ITERS,
     prob_opcode: float = DEFAULT_PROB_OPCODE,
     prob_operand: float = DEFAULT_PROB_OPERAND,
@@ -45,16 +46,28 @@ def optimize(
     padded_prgm = prgm.pad(max_size)
 
     last_prgm = padded_prgm
-    last_cost = cost(padded_prgm, test_cases, outputs, last_prgm)
+    last_cost, _last_eq = cost(padded_prgm, test_cases, outputs, last_prgm)
+
+    best_prgm = padded_prgm
+    best_cost = last_cost
 
     for _ in range(num_iters):
         candidate_prgm = mutate_program(
             last_prgm, prob_opcode, prob_operand, prob_swap, prob_insn, prob_insn_unused
         )
-        candidate_cost = cost(padded_prgm, test_cases, outputs, candidate_prgm)
+        candidate_cost, candidate_eq = cost(
+            padded_prgm, test_cases, outputs, candidate_prgm
+        )
+
+        if candidate_cost < best_cost and candidate_eq:
+            best_prgm = candidate_prgm
+            best_cost = candidate_cost
 
         if candidate_cost < last_cost - log(random()) / beta:
             last_prgm = candidate_prgm
             last_cost = candidate_cost
 
-    return last_prgm
+    print("last")
+    last_prgm.display()
+
+    return best_prgm