Adding ModCMA environment

.gitignore

@@ -16,4 +16,9 @@ tests/data/
 test/*.json
 
 # pycharm specific stuff
-.idea
+.idea
+
+
+venv
+settings.json
+*egg

dacbench/abstract_benchmark.py

@@ -86,6 +86,8 @@ def save_config(self, path):
                             and -np.inf not in conf[k][i]
                         ):
                             conf[k][i] = list(map(int, conf[k][i]))
+                elif isinstance(conf[k],np.ndarray):                    
+                    conf[k] = conf[k].tolist()
 
         conf["wrappers"] = self.jsonify_wrappers()
 
@@ -201,9 +203,12 @@ def read_config_file(self, path):
         if "observation_space_type" in self.config:
             # Types have to be numpy dtype (for gym spaces)s
             if type(self.config["observation_space_type"]) == str:
-                typestring = self.config["observation_space_type"].split(" ")[1][:-2]
-                typestring = typestring.split(".")[1]
-                self.config["observation_space_type"] = getattr(np, typestring)
+                if self.config["observation_space_type"] == "None":
+                    self.config["observation_space_type"] = None
+                else:
+                    typestring = self.config["observation_space_type"].split(" ")[1][:-2]
+                    typestring = typestring.split(".")[1]
+                    self.config["observation_space_type"] = getattr(np, typestring)
         if "observation_space" in self.config:
             self.config["observation_space"] = self.list_to_space(
                 self.config["observation_space"]
@@ -307,18 +312,15 @@ class objdict(dict):
     """
     Modified dict to make config changes more flexible
     """
+
+    __getattr__ = dict.__getitem__
+    __setattr__ = dict.__setitem__
+    __delattr__ = dict.__delitem__
+
+    def copy(self):
+        return objdict(**super().copy())
 
-    def __getattr__(self, name):
-        if name in self:
-            return self[name]
-        else:
-            raise AttributeError("No such attribute: " + name)
 
-    def __setattr__(self, name, value):
-        self[name] = value
+
+
 
-    def __delattr__(self, name):
-        if name in self:
-            del self[name]
-        else:
-            raise AttributeError("No such attribute: " + name)

dacbench/additional_configs/onell/configs.py

@@ -0,0 +1,122 @@
+"""
+All configs for the (1+(lambda, lambda)) Genetic Algorithms
+To generate the .json config files in this folder, run: export_all_configs_to_json()
+"""
+
+from copy import copy
+from dacbench.abstract_benchmark import objdict, AbstractBenchmark
+import numpy as np
+import json
+
+INFO = {
+    "identifier": "onell",
+    "name": "OneLL-GA benchmark",
+    "reward": "",
+    "state_description": [""]
+}
+
+"""
+Setting 1a: 
+  Name: lbd_theory
+  Description: 
+      The simplest setting where we only tune lbd and assume that p=lambda/n and c=1/lambda
+      The optimal policy for OneMax is lambda = sqrt(n / (n-f(x))), where:
+          n is problem size
+          f(x) is current objective value
+  State space: n (int), f(x) (int)
+  Action space: lbd (int)
+"""
+onell_lbd_theory = objdict(
+    {
+        "name": "lbd_theory",
+        "action_space_class": "Box",
+        "action_space_args": [np.array([1]), np.array([np.inf])],
+        "action_description": "lbd",
+        "observation_space_class": "Box", 
+        "observation_space_type": np.int32,       
+        "observation_space_args": [     
+            np.array([1, 0]),
+            np.array([np.inf, np.inf])   
+        ],
+        "observation_description": "n, f(x)",
+        "reward_range": [-np.inf, np.inf],   # the true reward range is instance dependent
+        "cutoff": 1e9,  # we don't really use this, 
+                        # the real cutoff is in instance_set_path and is instance dependent
+        "include_xprime": True, # if True, xprime is included in the selection after crossover phase
+        "count_different_inds_only": True, # if True, only count an evaluation of a child if it is different from both of its parents
+        "seed": 0,        
+        "problem": "OneMax",
+        "instance_set_path": "../instance_sets/onell/onemax_2000.csv",
+        "benchmark_info": INFO
+    }
+)
+
+# Setting 1b: 
+#   Name: lbd_onefifth
+#   Description: 
+#       Same as setting 1a but with slightly different state-space (n, delta f(x), lambda_{t-1})
+#       In this setting, the agent can learn the best policy for OneMax using the 1/5th rule:
+#           if delta f(x) = f(x_t) - f(x_{t-1}) <= 0: lambda_t = min{(3/2)^1/4 * lambda_{t-1}, n-1}
+#           otherwise: lambda_t = max{2/3 * lambda_{t-1}, 1}
+onell_lbd_onefifth = copy(onell_lbd_theory)
+onell_lbd_onefifth["name"] = "lbd_onefifth"
+onell_lbd_onefifth["observation_space_type"] = np.float32
+onell_lbd_onefifth["observation_space_args"] = [np.array([1, 0, 1]), np.array([np.inf, np.inf, np.inf])]
+onell_lbd_onefifth["observation_description"] = "n, delta f(x), lbd_{t-1}"
+
+
+# Setting 2: 
+#   Name: lbd_p_c
+#   Description: a more sophisticated setting where we tune lambda, p and c together
+#   State space: n (int), f(x) (int), delta f(x) (int), lambda_{t-1}, p_{t-1}, c_{t-1}
+#   action space: lambda (int), p (float), c (float)
+onell_lbd_p_c = copy(onell_lbd_theory)
+onell_lbd_p_c["name"] = "lbd_p_c"
+onell_lbd_p_c["observation_space_type"] = np.float32
+onell_lbd_p_c["observation_space_args"] = [np.array([1, 0, 0, 1, 0, 0]), np.array([np.inf, np.inf, np.inf, np.inf, 1, 1])]
+onell_lbd_p_c["action_space_args"] = [np.array([1, 0, 0]), np.array([np.inf, 1, 1])]
+onell_lbd_p_c["action_description"] = "lbd, p, c"
+onell_lbd_p_c["observation_description"] = "n, f(x), delta f(x), lbd_{t-1}, p_{t-1}, c_{t-1}"
+
+# Setting 3: 
+#   Name: lbd1_lbd2_p_c
+#   Description: a setting where we tune lambda1 (#mutated off-springs), lambda2 (#crossovered off-springs), p and c together
+#   State space: n (int), f(x) (int), delta f(x) (int), lambda1_{t-1}, lambda2_{t-1}, p_{t-1}, c_{t-1}
+#   action space: lambda1 (int), lambda2 (int), p (float), c (float)
+onell_lbd1_lbd2_p_c = copy(onell_lbd_theory)
+onell_lbd1_lbd2_p_c["name"] = "lbd1_lbd2_p_c"
+onell_lbd1_lbd2_p_c["observation_space_type"] = np.float32
+onell_lbd1_lbd2_p_c["observation_space_args"] = [np.array([1, 0, 0, 1, 1, 0, 0]), np.array([np.inf, np.inf, np.inf, np.inf, np.inf, 1, 1])]
+onell_lbd1_lbd2_p_c["action_space_args"] = [np.array([1, 1, 0, 0]), np.array([np.inf, np.inf, 1, 1])]
+onell_lbd1_lbd2_p_c["action_description"] = "lbd1, lbd2, p, c"
+onell_lbd1_lbd2_p_c["observation_description"] = "n, f(x), delta f(x), lbd1_{t-1}, lbd2_{t-1}, p_{t-1}, c_{t-1}"
+
+def config_to_json(config_name, json_file=None):
+    """
+    Write a config to json file
+
+    Parameters
+    ---------
+    config_name: str
+        accept values: one of the configs defined above
+    json_file: str  
+        output .json file name. If None, will be set as <config_name>.json
+    """
+    if json_file is None:
+        json_file = config_name + ".json"
+
+    with open(json_file, 'wt') as f:        
+        bench = AbstractBenchmark()
+        bench.config = globals()['onell_' + config_name]
+        bench.save_config(json_file)
+
+
+def export_all_configs_to_json(output_dir = './'):
+    """
+    Export all configs above to json files
+    """
+    for config_name in ['lbd_theory', 'lbd_onefifth', 'lbd_p_c', 'lbd1_lbd2_p_c']:
+        config_to_json(config_name, json_file=output_dir + '/' + config_name + '.json')
+
+
+#export_all_configs_to_json()

dacbench/additional_configs/onell/lbd1_lbd2_p_c.json

@@ -0,0 +1 @@
+{"name": "lbd1_lbd2_p_c", "action_space_class": "Box", "action_space_args": [[1, 1, 0, 0], [Infinity, Infinity, 1.0, 1.0]], "action_description": "lbd1, lbd2, p, c", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 0, 1, 1, 0, 0], [Infinity, Infinity, Infinity, Infinity, Infinity, 1.0, 1.0]], "observation_description": "n, f(x), delta f(x), lbd1_{t-1}, lbd2_{t-1}, p_{t-1}, c_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/lbd_onefifth.json

@@ -0,0 +1 @@
+{"name": "lbd_onefifth", "action_space_class": "Box", "action_space_args": [[1], [Infinity]], "action_description": "lbd", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 1], [Infinity, Infinity, Infinity]], "observation_description": "n, delta f(x), lbd_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/lbd_p_c.json

@@ -0,0 +1 @@
+{"name": "lbd_p_c", "action_space_class": "Box", "action_space_args": [[1, 0, 0], [Infinity, 1.0, 1.0]], "action_description": "lbd, p, c", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 0, 1, 0, 0], [Infinity, Infinity, Infinity, Infinity, 1.0, 1.0]], "observation_description": "n, f(x), delta f(x), lbd_{t-1}, p_{t-1}, c_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/lbd_theory.json

@@ -0,0 +1 @@
+{"name": "lbd_theory", "action_space_class": "Box", "action_space_args": [[1], [Infinity]], "action_description": "lbd", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.int32'>", "observation_space_args": [[1, 0], [Infinity, Infinity]], "observation_description": "n, f(x)", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/onell/configs.py

@@ -0,0 +1,122 @@
+"""
+All configs for the (1+(lambda, lambda)) Genetic Algorithms
+To generate the .json config files in this folder, run: export_all_configs_to_json()
+"""
+
+from copy import copy
+from dacbench.abstract_benchmark import objdict, AbstractBenchmark
+import numpy as np
+import json
+
+INFO = {
+    "identifier": "onell",
+    "name": "OneLL-GA benchmark",
+    "reward": "",
+    "state_description": [""]
+}
+
+"""
+Setting 1a: 
+  Name: lbd_theory
+  Description: 
+      The simplest setting where we only tune lbd and assume that p=lambda/n and c=1/lambda
+      The optimal policy for OneMax is lambda = sqrt(n / (n-f(x))), where:
+          n is problem size
+          f(x) is current objective value
+  State space: n (int), f(x) (int)
+  Action space: lbd (int)
+"""
+onell_lbd_theory = objdict(
+    {
+        "name": "lbd_theory",
+        "action_space_class": "Box",
+        "action_space_args": [np.array([1]), np.array([np.inf])],
+        "action_description": "lbd",
+        "observation_space_class": "Box", 
+        "observation_space_type": np.int32,       
+        "observation_space_args": [     
+            np.array([1, 0]),
+            np.array([np.inf, np.inf])   
+        ],
+        "observation_description": "n, f(x)",
+        "reward_range": [-np.inf, np.inf],   # the true reward range is instance dependent
+        "cutoff": 1e9,  # we don't really use this, 
+                        # the real cutoff is in instance_set_path and is instance dependent
+        "include_xprime": True, # if True, xprime is included in the selection after crossover phase
+        "count_different_inds_only": True, # if True, only count an evaluation of a child if it is different from both of its parents
+        "seed": 0,        
+        "problem": "OneMax",
+        "instance_set_path": "../instance_sets/onell/onemax_2000.csv",
+        "benchmark_info": INFO
+    }
+)
+
+# Setting 1b: 
+#   Name: lbd_onefifth
+#   Description: 
+#       Same as setting 1a but with slightly different state-space (n, delta f(x), lambda_{t-1})
+#       In this setting, the agent can learn the best policy for OneMax using the 1/5th rule:
+#           if delta f(x) = f(x_t) - f(x_{t-1}) <= 0: lambda_t = min{(3/2)^1/4 * lambda_{t-1}, n-1}
+#           otherwise: lambda_t = max{2/3 * lambda_{t-1}, 1}
+onell_lbd_onefifth = copy(onell_lbd_theory)
+onell_lbd_onefifth["name"] = "lbd_onefifth"
+onell_lbd_onefifth["observation_space_type"] = np.float32
+onell_lbd_onefifth["observation_space_args"] = [np.array([1, 0, 1]), np.array([np.inf, np.inf, np.inf])]
+onell_lbd_onefifth["observation_description"] = "n, delta f(x), lbd_{t-1}"
+
+
+# Setting 2: 
+#   Name: lbd_p_c
+#   Description: a more sophisticated setting where we tune lambda, p and c together
+#   State space: n (int), f(x) (int), delta f(x) (int), lambda_{t-1}, p_{t-1}, c_{t-1}
+#   action space: lambda (int), p (float), c (float)
+onell_lbd_p_c = copy(onell_lbd_theory)
+onell_lbd_p_c["name"] = "lbd_p_c"
+onell_lbd_p_c["observation_space_type"] = np.float32
+onell_lbd_p_c["observation_space_args"] = [np.array([1, 0, 0, 1, 0, 0]), np.array([np.inf, np.inf, np.inf, np.inf, 1, 1])]
+onell_lbd_p_c["action_space_args"] = [np.array([1, 0, 0]), np.array([np.inf, 1, 1])]
+onell_lbd_p_c["action_description"] = "lbd, p, c"
+onell_lbd_p_c["observation_description"] = "n, f(x), delta f(x), lbd_{t-1}, p_{t-1}, c_{t-1}"
+
+# Setting 3: 
+#   Name: lbd1_lbd2_p_c
+#   Description: a setting where we tune lambda1 (#mutated off-springs), lambda2 (#crossovered off-springs), p and c together
+#   State space: n (int), f(x) (int), delta f(x) (int), lambda1_{t-1}, lambda2_{t-1}, p_{t-1}, c_{t-1}
+#   action space: lambda1 (int), lambda2 (int), p (float), c (float)
+onell_lbd1_lbd2_p_c = copy(onell_lbd_theory)
+onell_lbd1_lbd2_p_c["name"] = "lbd1_lbd2_p_c"
+onell_lbd1_lbd2_p_c["observation_space_type"] = np.float32
+onell_lbd1_lbd2_p_c["observation_space_args"] = [np.array([1, 0, 0, 1, 1, 0, 0]), np.array([np.inf, np.inf, np.inf, np.inf, np.inf, 1, 1])]
+onell_lbd1_lbd2_p_c["action_space_args"] = [np.array([1, 1, 0, 0]), np.array([np.inf, np.inf, 1, 1])]
+onell_lbd1_lbd2_p_c["action_description"] = "lbd1, lbd2, p, c"
+onell_lbd1_lbd2_p_c["observation_description"] = "n, f(x), delta f(x), lbd1_{t-1}, lbd2_{t-1}, p_{t-1}, c_{t-1}"
+
+def config_to_json(config_name, json_file=None):
+    """
+    Write a config to json file
+
+    Parameters
+    ---------
+    config_name: str
+        accept values: one of the configs defined above
+    json_file: str  
+        output .json file name. If None, will be set as <config_name>.json
+    """
+    if json_file is None:
+        json_file = config_name + ".json"
+
+    with open(json_file, 'wt') as f:        
+        bench = AbstractBenchmark()
+        bench.config = globals()['onell_' + config_name]
+        bench.save_config(json_file)
+
+
+def export_all_configs_to_json(output_dir = './'):
+    """
+    Export all configs above to json files
+    """
+    for config_name in ['lbd_theory', 'lbd_onefifth', 'lbd_p_c', 'lbd1_lbd2_p_c']:
+        config_to_json(config_name, json_file=output_dir + '/' + config_name + '.json')
+
+
+#export_all_configs_to_json()

dacbench/additional_configs/onell/onell/lbd1_lbd2_p_c.json

@@ -0,0 +1 @@
+{"name": "lbd1_lbd2_p_c", "action_space_class": "Box", "action_space_args": [[1, 1, 0, 0], [Infinity, Infinity, 1.0, 1.0]], "action_description": "lbd1, lbd2, p, c", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 0, 1, 1, 0, 0], [Infinity, Infinity, Infinity, Infinity, Infinity, 1.0, 1.0]], "observation_description": "n, f(x), delta f(x), lbd1_{t-1}, lbd2_{t-1}, p_{t-1}, c_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/onell/lbd_onefifth.json

@@ -0,0 +1 @@
+{"name": "lbd_onefifth", "action_space_class": "Box", "action_space_args": [[1], [Infinity]], "action_description": "lbd", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 1], [Infinity, Infinity, Infinity]], "observation_description": "n, delta f(x), lbd_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/onell/lbd_p_c.json

@@ -0,0 +1 @@
+{"name": "lbd_p_c", "action_space_class": "Box", "action_space_args": [[1, 0, 0], [Infinity, 1.0, 1.0]], "action_description": "lbd, p, c", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.float32'>", "observation_space_args": [[1, 0, 0, 1, 0, 0], [Infinity, Infinity, Infinity, Infinity, 1.0, 1.0]], "observation_description": "n, f(x), delta f(x), lbd_{t-1}, p_{t-1}, c_{t-1}", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/additional_configs/onell/onell/lbd_theory.json

@@ -0,0 +1 @@
+{"name": "lbd_theory", "action_space_class": "Box", "action_space_args": [[1], [Infinity]], "action_description": "lbd", "observation_space_class": "Box", "observation_space_type": "<class 'numpy.int32'>", "observation_space_args": [[1, 0], [Infinity, Infinity]], "observation_description": "n, f(x)", "reward_range": [-Infinity, Infinity], "cutoff": 1000000000.0, "include_xprime": true, "count_different_inds_only": true, "seed": 0, "problem": "OneMax", "instance_set_path": "../instance_sets/onell/onemax_2000.csv", "benchmark_info": {"identifier": "onell", "name": "OneLL-GA benchmark", "reward": "", "state_description": [""]}, "wrappers": []}

dacbench/benchmarks/__init__.py

@@ -4,6 +4,7 @@
 from dacbench.benchmarks.cma_benchmark import CMAESBenchmark
 from dacbench.benchmarks.modea_benchmark import ModeaBenchmark
 from dacbench.benchmarks.sgd_benchmark import SGDBenchmark
+from dacbench.benchmarks.modcma_benchmark import ModCMABenchmark
 
 __all__ = [
     "LubyBenchmark",
@@ -12,4 +13,5 @@
     "CMAESBenchmark",
     "ModeaBenchmark",
     "SGDBenchmark",
+    "ModCMABenchmark"
 ]