The examples weren't being properly tested in the last commit. When we fixed
that a lot of bugs in the new implementation of environment and agent were
found, which accounts for most of these changes.
The main difference is the mechanism to load simulations from a configuration
file. For that to work, we had to rework our module loading code in
`serialization` and add a `source_file` attribute to configurations (and
simulations, for that matter).
All test pass, except for the TestConfig suite, which is not too critical as the
plan for this version onwards is to avoid configuration as much as possible.
* Removed old/unnecessary models
* Added a `simulation.{iter_}from_py` method to load simulations from python
files
* Changed tests of examples to run programmatic simulations
* Fixed programmatic examples
Treating time and conditions as the same entity was getting confusing, and it
added a lot of unnecessary abstraction in a critical part (the scheduler).
The scheduling queue now has the time as a floating number (faster), the agent
id (for ties) and the condition, as well as the agent. The first three
elements (time, id, condition) can be considered as the "key" for the event.
To allow for agent execution to be "randomized" within every step, a new
parameter has been added to the scheduler, which makes it add a random number to
the key in order to change the ordering.
`EventedAgent.received` now checks the messages before returning control to the
user by default.
* Removed references to `set_state`
* Split some functionality from `agents` into separate files (`fsm` and
`network_agents`)
* Rename `neighboring_agents` to `neighbors`
* Delete some spurious functions
Documentation needs some improvement
The API has been simplified to only allow for ONE topology per
NetworkEnvironment.
This covers the main use case, and simplifies the code.
All tests pass but some features are still missing/unclear:
- Mesa agents do not have a `state`, so their "metrics" don't get stored. I will
probably refactor this to remove some magic in this regard. This should get rid
of the `_state` dictionary and the setitem/getitem magic.
- Simulation is still different from a runner. So far only Agent and
Environment/Model have been updated.
See CHANGELOG.md for a full list of changes
* Removed nxsim
* Refactored `agents.NetworkAgent` and `agents.BaseAgent`
* Refactored exporters
* Added stats to history
Multithreading needs pickling to work.
Pickling/unpickling didn't work in some situations, like when the
environment_agents parameter was left blank.
This was due to two reasons:
1) agents and history didn't have a setstate method, and some of their
attributes cannot be pickled (generators, sqlite connection)
2) the environment was adding generators (agents) to its state.
This fixes the situation by restricting the keys that the environment exports
when it pickles, and by adding the set/getstate methods in agents.
The resulting pickles should contain enough information to inspect
them (history, state values, etc), but very limited.
* Upgrade to python3.7 and pandas 0.3.4 because pandas has dropped support for
python 3.4 -> There are some API changes in pandas, and I've update the code
accordingly.
* Set pytest as the default test runner
Now the environment does not deal with history directly, it delegates it to a
specific class. The analysis also uses history instances instead of either
using the database directly or creating a proxy environment.
This should make it easier to change the implementation in the future.
In fact, the change was motivated by the large size of the csv files in previous
versions. This new implementation only stores results in deltas, and it fills
any necessary values when needed.
