mirror of https://github.com/gsi-upm/soil
You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
673 lines
19 KiB
Python
673 lines
19 KiB
Python
from __future__ import annotations
|
|
|
|
import logging
|
|
from collections import OrderedDict, defaultdict
|
|
from collections.abc import MutableMapping, Mapping, Set
|
|
from abc import ABCMeta
|
|
from copy import deepcopy, copy
|
|
from functools import partial, wraps
|
|
from itertools import islice, chain
|
|
import inspect
|
|
import types
|
|
import textwrap
|
|
import networkx as nx
|
|
import warnings
|
|
import sys
|
|
|
|
from typing import Any
|
|
|
|
from mesa import Agent as MesaAgent, Model
|
|
from typing import Dict, List
|
|
|
|
from .. import serialization, network, utils, time, config
|
|
|
|
|
|
IGNORED_FIELDS = ("model", "logger")
|
|
|
|
|
|
class MetaAgent(ABCMeta):
|
|
def __new__(mcls, name, bases, namespace):
|
|
defaults = {}
|
|
|
|
# Re-use defaults from inherited classes
|
|
for i in bases:
|
|
if isinstance(i, MetaAgent):
|
|
defaults.update(i._defaults)
|
|
|
|
new_nmspc = {
|
|
"_defaults": defaults,
|
|
"_last_return": None,
|
|
"_last_except": None,
|
|
}
|
|
|
|
for attr, func in namespace.items():
|
|
if attr == "step" and inspect.isgeneratorfunction(func):
|
|
orig_func = func
|
|
new_nmspc["_coroutine"] = None
|
|
|
|
@wraps(func)
|
|
def func(self):
|
|
while True:
|
|
if not self._coroutine:
|
|
self._coroutine = orig_func(self)
|
|
try:
|
|
if self._last_except:
|
|
return self._coroutine.throw(self._last_except)
|
|
else:
|
|
return self._coroutine.send(self._last_return)
|
|
except StopIteration as ex:
|
|
self._coroutine = None
|
|
return ex.value
|
|
finally:
|
|
self._last_return = None
|
|
self._last_except = None
|
|
|
|
func.id = name or func.__name__
|
|
func.is_default = False
|
|
new_nmspc[attr] = func
|
|
elif (
|
|
isinstance(func, types.FunctionType)
|
|
or isinstance(func, property)
|
|
or isinstance(func, classmethod)
|
|
or attr[0] == "_"
|
|
):
|
|
new_nmspc[attr] = func
|
|
elif attr == "defaults":
|
|
defaults.update(func)
|
|
else:
|
|
defaults[attr] = copy(func)
|
|
|
|
return super().__new__(mcls=mcls, name=name, bases=bases, namespace=new_nmspc)
|
|
|
|
|
|
class BaseAgent(MesaAgent, MutableMapping, metaclass=MetaAgent):
|
|
"""
|
|
A special type of Mesa Agent that:
|
|
|
|
* Can be used as a dictionary to access its state.
|
|
* Has logging built-in
|
|
* Can be given default arguments through a defaults class attribute,
|
|
which will be used on construction to initialize each agent's state
|
|
|
|
Any attribute that is not preceded by an underscore (`_`) will also be added to its state.
|
|
"""
|
|
|
|
def __init__(self, unique_id, model, name=None, init=True, interval=None, **kwargs):
|
|
assert isinstance(unique_id, int)
|
|
super().__init__(unique_id=unique_id, model=model)
|
|
|
|
self.name = (
|
|
str(name) if name else "{}[{}]".format(type(self).__name__, self.unique_id)
|
|
)
|
|
|
|
self.alive = True
|
|
|
|
self.interval = interval or self.get("interval", 1)
|
|
logger = utils.logger.getChild(getattr(self.model, "id", self.model)).getChild(
|
|
self.name
|
|
)
|
|
self.logger = logging.LoggerAdapter(logger, {"agent_name": self.name})
|
|
|
|
if hasattr(self, "level"):
|
|
self.logger.setLevel(self.level)
|
|
|
|
for (k, v) in self._defaults.items():
|
|
if not hasattr(self, k) or getattr(self, k) is None:
|
|
setattr(self, k, deepcopy(v))
|
|
|
|
for (k, v) in kwargs.items():
|
|
|
|
setattr(self, k, v)
|
|
if init:
|
|
self.init()
|
|
|
|
def init(self):
|
|
pass
|
|
|
|
def __hash__(self):
|
|
return hash(self.unique_id)
|
|
|
|
def prob(self, probability):
|
|
return prob(probability, self.model.random)
|
|
|
|
@classmethod
|
|
def w(cls, **kwargs):
|
|
return custom(cls, **kwargs)
|
|
|
|
# TODO: refactor to clean up mesa compatibility
|
|
@property
|
|
def id(self):
|
|
msg = "This attribute is deprecated. Use `unique_id` instead"
|
|
warnings.warn(msg, DeprecationWarning)
|
|
print(msg, file=sys.stderr)
|
|
return self.unique_id
|
|
|
|
@classmethod
|
|
def from_dict(cls, model, attrs, warn_extra=True):
|
|
ignored = {}
|
|
args = {}
|
|
for k, v in attrs.items():
|
|
if k in inspect.signature(cls).parameters:
|
|
args[k] = v
|
|
else:
|
|
ignored[k] = v
|
|
if ignored and warn_extra:
|
|
utils.logger.info(
|
|
f"Ignoring the following arguments for agent class { agent_class.__name__ }: { ignored }"
|
|
)
|
|
return cls(model=model, **args)
|
|
|
|
def __getitem__(self, key):
|
|
try:
|
|
return getattr(self, key)
|
|
except AttributeError:
|
|
raise KeyError(f"key {key} not found in agent")
|
|
|
|
def __delitem__(self, key):
|
|
return delattr(self, key)
|
|
|
|
def __contains__(self, key):
|
|
return hasattr(self, key)
|
|
|
|
def __setitem__(self, key, value):
|
|
setattr(self, key, value)
|
|
|
|
def __len__(self):
|
|
return sum(1 for n in self.keys())
|
|
|
|
def __iter__(self):
|
|
return self.items()
|
|
|
|
def keys(self):
|
|
return (k for k in self.__dict__ if k[0] != "_" and k not in IGNORED_FIELDS)
|
|
|
|
def items(self, keys=None, skip=None):
|
|
keys = keys if keys is not None else self.keys()
|
|
it = ((k, self.get(k, None)) for k in keys)
|
|
if skip:
|
|
return filter(lambda x: x[0] not in skip, it)
|
|
return it
|
|
|
|
def get(self, key, default=None):
|
|
if key in self:
|
|
return self[key]
|
|
elif key in self.model:
|
|
return self.model[key]
|
|
return default
|
|
|
|
@property
|
|
def now(self):
|
|
try:
|
|
return self.model.now
|
|
except AttributeError:
|
|
# No environment
|
|
return None
|
|
|
|
def die(self, msg=None):
|
|
if msg:
|
|
self.info("Agent dying:", msg)
|
|
self.debug(f"agent dying")
|
|
self.alive = False
|
|
try:
|
|
self.model.schedule.remove(self)
|
|
except KeyError:
|
|
pass
|
|
return time.NEVER
|
|
|
|
def step(self):
|
|
raise NotImplementedError("Agent must implement step method")
|
|
|
|
def _check_alive(self):
|
|
if not self.alive:
|
|
raise time.DeadAgent(self.unique_id)
|
|
|
|
def log(self, *message, level=logging.INFO, **kwargs):
|
|
if not self.logger.isEnabledFor(level):
|
|
return
|
|
message = " ".join(str(i) for i in message)
|
|
message = "[@{:>4}]\t{:>10}: {}".format(self.now, repr(self), message)
|
|
for k, v in kwargs:
|
|
message += " {k}={v} ".format(k, v)
|
|
extra = {}
|
|
extra["now"] = self.now
|
|
extra["unique_id"] = self.unique_id
|
|
extra["agent_name"] = self.name
|
|
return self.logger.log(level, message, extra=extra)
|
|
|
|
def debug(self, *args, **kwargs):
|
|
return self.log(*args, level=logging.DEBUG, **kwargs)
|
|
|
|
def info(self, *args, **kwargs):
|
|
return self.log(*args, level=logging.INFO, **kwargs)
|
|
|
|
def count_agents(self, **kwargs):
|
|
return len(list(self.get_agents(**kwargs)))
|
|
|
|
def get_agents(self, *args, **kwargs):
|
|
it = self.iter_agents(*args, **kwargs)
|
|
return list(it)
|
|
|
|
def iter_agents(self, *args, **kwargs):
|
|
yield from filter_agents(self.model.schedule._agents, *args, **kwargs)
|
|
|
|
def __str__(self):
|
|
return self.to_str()
|
|
|
|
def to_str(self, keys=None, skip=None, pretty=False):
|
|
content = dict(self.items(keys=keys))
|
|
if pretty and content:
|
|
d = content
|
|
content = "\n"
|
|
for k, v in d.items():
|
|
content += f"- {k}: {v}\n"
|
|
content = textwrap.indent(content, " ")
|
|
return f"{repr(self)}{content}"
|
|
|
|
def __repr__(self):
|
|
return f"{self.__class__.__name__}({self.unique_id})"
|
|
|
|
|
|
def prob(prob, random):
|
|
"""
|
|
A true/False uniform distribution with a given probability.
|
|
To be used like this:
|
|
|
|
.. code-block:: python
|
|
|
|
if prob(0.3):
|
|
do_something()
|
|
|
|
"""
|
|
r = random.random()
|
|
return r < prob
|
|
|
|
|
|
def calculate_distribution(network_agents=None, agent_class=None):
|
|
"""
|
|
Calculate the threshold values (thresholds for a uniform distribution)
|
|
of an agent distribution given the weights of each agent type.
|
|
|
|
The input has this form: ::
|
|
|
|
[
|
|
{'agent_class': 'agent_class_1',
|
|
'weight': 0.2,
|
|
'state': {
|
|
'id': 0
|
|
}
|
|
},
|
|
{'agent_class': 'agent_class_2',
|
|
'weight': 0.8,
|
|
'state': {
|
|
'id': 1
|
|
}
|
|
}
|
|
]
|
|
|
|
In this example, 20% of the nodes will be marked as type
|
|
'agent_class_1'.
|
|
"""
|
|
if network_agents:
|
|
network_agents = [
|
|
deepcopy(agent) for agent in network_agents if not hasattr(agent, "id")
|
|
]
|
|
elif agent_class:
|
|
network_agents = [{"agent_class": agent_class}]
|
|
else:
|
|
raise ValueError("Specify a distribution or a default agent type")
|
|
|
|
# Fix missing weights and incompatible types
|
|
for x in network_agents:
|
|
x["weight"] = float(x.get("weight", 1))
|
|
|
|
# Calculate the thresholds
|
|
total = sum(x["weight"] for x in network_agents)
|
|
acc = 0
|
|
for v in network_agents:
|
|
if "ids" in v:
|
|
continue
|
|
upper = acc + (v["weight"] / total)
|
|
v["threshold"] = [acc, upper]
|
|
acc = upper
|
|
return network_agents
|
|
|
|
|
|
def _serialize_type(agent_class, known_modules=[], **kwargs):
|
|
if isinstance(agent_class, str):
|
|
return agent_class
|
|
known_modules += ["soil.agents"]
|
|
return serialization.serialize(agent_class, known_modules=known_modules, **kwargs)[
|
|
1
|
|
] # Get the name of the class
|
|
|
|
|
|
def _deserialize_type(agent_class, known_modules=[]):
|
|
if not isinstance(agent_class, str):
|
|
return agent_class
|
|
known = known_modules + ["soil.agents", "soil.agents.custom"]
|
|
agent_class = serialization.deserializer(agent_class, known_modules=known)
|
|
return agent_class
|
|
|
|
|
|
class AgentView(Mapping, Set):
|
|
"""A lazy-loaded list of agents."""
|
|
|
|
__slots__ = ("_agents",)
|
|
|
|
def __init__(self, agents):
|
|
self._agents = agents
|
|
|
|
def __getstate__(self):
|
|
return {"_agents": self._agents}
|
|
|
|
def __setstate__(self, state):
|
|
self._agents = state["_agents"]
|
|
|
|
# Mapping methods
|
|
def __len__(self):
|
|
return len(self._agents)
|
|
|
|
def __iter__(self):
|
|
yield from self._agents.values()
|
|
|
|
def __getitem__(self, agent_id):
|
|
if isinstance(agent_id, slice):
|
|
raise ValueError(f"Slicing is not supported")
|
|
if agent_id in self._agents:
|
|
return self._agents[agent_id]
|
|
raise ValueError(f"Agent {agent_id} not found")
|
|
|
|
def filter(self, *args, **kwargs):
|
|
yield from filter_agents(self._agents, *args, **kwargs)
|
|
|
|
def one(self, *args, **kwargs):
|
|
return next(filter_agents(self._agents, *args, **kwargs))
|
|
|
|
def __call__(self, *args, **kwargs):
|
|
return list(self.filter(*args, **kwargs))
|
|
|
|
def __contains__(self, agent_id):
|
|
return agent_id in self._agents
|
|
|
|
def __str__(self):
|
|
return str(list(unique_id for unique_id in self.keys()))
|
|
|
|
def __repr__(self):
|
|
return f"{self.__class__.__name__}({self})"
|
|
|
|
|
|
def filter_agents(
|
|
agents: dict,
|
|
*id_args,
|
|
unique_id=None,
|
|
state_id=None,
|
|
agent_class=None,
|
|
ignore=None,
|
|
state=None,
|
|
limit=None,
|
|
**kwargs,
|
|
):
|
|
"""
|
|
Filter agents given as a dict, by the criteria given as arguments (e.g., certain type or state id).
|
|
"""
|
|
assert isinstance(agents, dict)
|
|
|
|
ids = []
|
|
|
|
if unique_id is not None:
|
|
if isinstance(unique_id, list):
|
|
ids += unique_id
|
|
else:
|
|
ids.append(unique_id)
|
|
|
|
if id_args:
|
|
ids += id_args
|
|
|
|
if ids:
|
|
f = (agents[aid] for aid in ids if aid in agents)
|
|
else:
|
|
f = agents.values()
|
|
|
|
if state_id is not None and not isinstance(state_id, (tuple, list)):
|
|
state_id = tuple([state_id])
|
|
|
|
if agent_class is not None:
|
|
agent_class = _deserialize_type(agent_class)
|
|
try:
|
|
agent_class = tuple(agent_class)
|
|
except TypeError:
|
|
agent_class = tuple([agent_class])
|
|
|
|
if ignore:
|
|
f = filter(lambda x: x not in ignore, f)
|
|
|
|
if state_id is not None:
|
|
f = filter(lambda agent: agent.get("state_id", None) in state_id, f)
|
|
|
|
if agent_class is not None:
|
|
f = filter(lambda agent: isinstance(agent, agent_class), f)
|
|
|
|
state = state or dict()
|
|
state.update(kwargs)
|
|
|
|
for k, v in state.items():
|
|
f = filter(lambda agent: getattr(agent, k, None) == v, f)
|
|
|
|
if limit is not None:
|
|
f = islice(f, limit)
|
|
|
|
yield from f
|
|
|
|
|
|
def from_config(
|
|
cfg: config.AgentConfig, random, topology: nx.Graph = None
|
|
) -> List[Dict[str, Any]]:
|
|
"""
|
|
This function turns an agentconfig into a list of individual "agent specifications", which are just a dictionary
|
|
with the parameters that the environment will use to construct each agent.
|
|
|
|
This function does NOT return a list of agents, mostly because some attributes to the agent are not known at the
|
|
time of calling this function, such as `unique_id`.
|
|
"""
|
|
default = cfg or config.AgentConfig()
|
|
if not isinstance(cfg, config.AgentConfig):
|
|
cfg = config.AgentConfig(**cfg)
|
|
|
|
agents = []
|
|
|
|
assigned_total = 0
|
|
assigned_network = 0
|
|
|
|
if cfg.fixed is not None:
|
|
agents, assigned_total, assigned_network = _from_fixed(
|
|
cfg.fixed, topology=cfg.topology, default=cfg
|
|
)
|
|
|
|
n = cfg.n
|
|
|
|
if cfg.distribution:
|
|
topo_size = len(topology) if topology else 0
|
|
|
|
networked = []
|
|
total = []
|
|
|
|
for d in cfg.distribution:
|
|
if d.strategy == config.Strategy.topology:
|
|
topo = d.topology if ("topology" in d.__fields_set__) else cfg.topology
|
|
if not topo:
|
|
raise ValueError(
|
|
'The "topology" strategy only works if the topology parameter is set to True'
|
|
)
|
|
if not topo_size:
|
|
raise ValueError(
|
|
f"Topology does not have enough free nodes to assign one to the agent"
|
|
)
|
|
|
|
networked.append(d)
|
|
|
|
if d.strategy == config.Strategy.total:
|
|
if not cfg.n:
|
|
raise ValueError(
|
|
'Cannot use the "total" strategy without providing the total number of agents'
|
|
)
|
|
total.append(d)
|
|
|
|
if networked:
|
|
new_agents = _from_distro(
|
|
networked,
|
|
n=topo_size - assigned_network,
|
|
topology=topo,
|
|
default=cfg,
|
|
random=random,
|
|
)
|
|
assigned_total += len(new_agents)
|
|
assigned_network += len(new_agents)
|
|
agents += new_agents
|
|
|
|
if total:
|
|
remaining = n - assigned_total
|
|
agents += _from_distro(total, n=remaining, default=cfg, random=random)
|
|
|
|
if assigned_network < topo_size:
|
|
utils.logger.warn(
|
|
f"The total number of agents does not match the total number of nodes in "
|
|
"every topology. This may be due to a definition error: assigned: "
|
|
f"{ assigned } total size: { topo_size }"
|
|
)
|
|
|
|
return agents
|
|
|
|
|
|
def _from_fixed(
|
|
lst: List[config.FixedAgentConfig],
|
|
topology: bool,
|
|
default: config.SingleAgentConfig,
|
|
) -> List[Dict[str, Any]]:
|
|
agents = []
|
|
|
|
counts_total = 0
|
|
counts_network = 0
|
|
|
|
for fixed in lst:
|
|
agent = {}
|
|
if default:
|
|
agent = default.state.copy()
|
|
agent.update(fixed.state)
|
|
cls = serialization.deserialize(
|
|
fixed.agent_class or (default and default.agent_class)
|
|
)
|
|
agent["agent_class"] = cls
|
|
topo = (
|
|
fixed.topology
|
|
if ("topology" in fixed.__fields_set__)
|
|
else topology or default.topology
|
|
)
|
|
|
|
if topo:
|
|
agent["topology"] = True
|
|
counts_network += 1
|
|
if not fixed.hidden:
|
|
counts_total += 1
|
|
agents.append(agent)
|
|
|
|
return agents, counts_total, counts_network
|
|
|
|
|
|
def _from_distro(
|
|
distro: List[config.AgentDistro],
|
|
n: int,
|
|
default: config.SingleAgentConfig,
|
|
random,
|
|
topology: str = None
|
|
) -> List[Dict[str, Any]]:
|
|
|
|
agents = []
|
|
|
|
if n is None:
|
|
if any(lambda dist: dist.n is None, distro):
|
|
raise ValueError(
|
|
"You must provide a total number of agents, or the number of each type"
|
|
)
|
|
n = sum(dist.n for dist in distro)
|
|
|
|
weights = list(dist.weight if dist.weight is not None else 1 for dist in distro)
|
|
minw = min(weights)
|
|
norm = list(weight / minw for weight in weights)
|
|
total = sum(norm)
|
|
chunk = n // total
|
|
|
|
# random.choices would be enough to get a weighted distribution. But it can vary a lot for smaller k
|
|
# So instead we calculate our own distribution to make sure the actual ratios are close to what we would expect
|
|
|
|
# Calculate how many times each has to appear
|
|
indices = list(
|
|
chain.from_iterable([idx] * int(n * chunk) for (idx, n) in enumerate(norm))
|
|
)
|
|
|
|
# Complete with random agents following the original weight distribution
|
|
if len(indices) < n:
|
|
indices += random.choices(
|
|
list(range(len(distro))),
|
|
weights=[d.weight for d in distro],
|
|
k=n - len(indices),
|
|
)
|
|
|
|
# Deserialize classes for efficiency
|
|
classes = list(
|
|
serialization.deserialize(i.agent_class or default.agent_class) for i in distro
|
|
)
|
|
|
|
# Add them in random order
|
|
random.shuffle(indices)
|
|
|
|
for idx in indices:
|
|
d = distro[idx]
|
|
agent = d.state.copy()
|
|
cls = classes[idx]
|
|
agent["agent_class"] = cls
|
|
if default:
|
|
agent.update(default.state)
|
|
topology = (
|
|
d.topology
|
|
if ("topology" in d.__fields_set__)
|
|
else topology or default.topology
|
|
)
|
|
if topology:
|
|
agent["topology"] = topology
|
|
agents.append(agent)
|
|
|
|
return agents
|
|
|
|
|
|
from .network_agents import *
|
|
from .fsm import *
|
|
from .evented import *
|
|
from typing import Optional
|
|
|
|
|
|
class Agent(NetworkAgent, FSM, EventedAgent):
|
|
"""Default agent class, has both network and event capabilities"""
|
|
|
|
|
|
from ..environment import NetworkEnvironment
|
|
|
|
|
|
from .BassModel import *
|
|
from .IndependentCascadeModel import *
|
|
from .SISaModel import *
|
|
from .CounterModel import *
|
|
|
|
|
|
try:
|
|
import scipy
|
|
from .Geo import Geo
|
|
except ImportError:
|
|
import sys
|
|
|
|
print("Could not load the Geo Agent, scipy is not installed", file=sys.stderr)
|
|
|
|
|
|
def custom(cls, **kwargs):
|
|
"""Create a new class from a template class and keyword arguments"""
|
|
return type(cls.__name__, (cls,), kwargs)
|