evolutionary-policy-optimization 0.0.6__py3-none-any.whl → 0.0.9__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -335,6 +335,25 @@ class Agent(Module):
     ):
         raise NotImplementedError
 
+# criteria for running genetic algorithm
+
+class ShouldRunGeneticAlgorithm(Module):
+    def __init__(
+        self,
+        gamma = 1.5 # not sure what the value is
+    ):
+        super().__init__()
+        self.gamma = gamma
+
+    def forward(self, fitnesses):
+        # equation (3)
+
+        # max(fitness) - min(fitness) > gamma * median(fitness)
+        # however, this equation does not make much sense to me if fitness increases unbounded
+        # just let it be customizable, and offer a variant where mean and variance is over some threshold (could account for skew too)
+
+        return (fitnesses.amax() - fitnesses.amin()) > (self.gamma * torch.median(fitnesses))
+
 # classes
 
 class LatentGenePool(Module):
@@ -350,6 +369,8 @@ class LatentGenePool(Module):
         frac_elitism = 0.1,              # frac of population to preserve from being noised
         mutation_strength = 1.,          # factor to multiply to gaussian noise as mutation to latents
         net: MLP | Module | dict | None = None,
+        should_run_genetic_algorithm: Module | None = None, # eq (3) in paper
+        default_should_run_ga_gamma = 1.5
     ):
         super().__init__()
 
@@ -400,6 +421,11 @@ class LatentGenePool(Module):
 
         self.net = net
 
+        if not exists(should_run_genetic_algorithm):
+            should_run_genetic_algorithm = ShouldRunGeneticAlgorithm(gamma = default_should_run_ga_gamma)
+
+        self.should_run_genetic_algorithm = should_run_genetic_algorithm
+
     @torch.no_grad()
     # non-gradient optimization, at least, not on the individual level (taken care of by rl component)
     def genetic_algorithm_step(
@@ -412,6 +438,10 @@ class LatentGenePool(Module):
         g - gene dimension
         n - number of genes per individual
         """
+
+        if not self.should_run_genetic_algorithm(fitness):
+            return
+
         assert self.num_latents > 1
 
         genes = self.latents # the latents are the genes

{evolutionary_policy_optimization-0.0.6.dist-info → evolutionary_policy_optimization-0.0.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.6
+Version: 0.0.9
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization

evolutionary_policy_optimization-0.0.9.dist-info/RECORD

@@ -0,0 +1,7 @@
+evolutionary_policy_optimization/__init__.py,sha256=Nu-_OMA8abe4AiW9Gw6MvbZH0JZpMHMqjeXmkC9-7UU,81
+evolutionary_policy_optimization/epo.py,sha256=eiOJg0J14miB3ZWpcTD1dMC7M6abxtVaMD_Oxza0cYI,15880
+evolutionary_policy_optimization/experimental.py,sha256=ktBKxRF27Qsj7WIgBpYlWXqMVxO9zOx2oD1JuDYRAwM,548
+evolutionary_policy_optimization-0.0.9.dist-info/METADATA,sha256=MT4_JXsUQCrcBWB-0m9uJZHYtGnSFMbQzclZ32HZKnQ,4460
+evolutionary_policy_optimization-0.0.9.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.9.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.9.dist-info/RECORD,,

evolutionary_policy_optimization-0.0.6.dist-info/RECORD

@@ -1,7 +0,0 @@
-evolutionary_policy_optimization/__init__.py,sha256=Nu-_OMA8abe4AiW9Gw6MvbZH0JZpMHMqjeXmkC9-7UU,81
-evolutionary_policy_optimization/epo.py,sha256=vXkwsQE0CNEUPpguZP-XXsuDyIBN-bS3xDJDXpYlTHM,14772
-evolutionary_policy_optimization/experimental.py,sha256=ktBKxRF27Qsj7WIgBpYlWXqMVxO9zOx2oD1JuDYRAwM,548
-evolutionary_policy_optimization-0.0.6.dist-info/METADATA,sha256=M_0SbTqdifHQ_R9LWIe7ZfHMXgCiFDJ0sDpD29ctiNk,4460
-evolutionary_policy_optimization-0.0.6.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.6.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.6.dist-info/RECORD,,