Directed search with the Exploratory Modeling workbench

This is the third blog in a series showcasing the functionality of the Exploratory Modeling workbench. Exploratory modeling entails investigating the way in which uncertainty and/or policy levers map to outcomes. To investigate these mappings, we can either use sampling based strategies (open exploration) or optimization based strategies (directed search) In the first blog, I gave a general overview of the workbench and showed briefly how both investigation strategies can be done. In the second blog, I demonstrated the use of the workbench for open exploration in substantial more detail. In this third blog, I will demonstrate in more detail how to use the workbench for directed search. Like in the previous two blog post, I will use the DPS version of the lake problem.

For optimization, the workbench relies on platypus. You can easily install the latest version of platypus from github using pip

pip install git+https://github.com/Project-Platypus/Platypus.git

By default, the workbench will use epsilon NSGA2, but all the other algorithms available within platypus can be used as well.

Within the workbench, optimization can be used in three ways:
* Search over decision levers for a reference scenario
* Robust search: search over decision levers for a set of scenarios
* worst case discovery: search over uncertainties for a reference policy

The search over decision levers or over uncertainties relies on the specification of the direction for each outcome of interest defined on the model. It is only possible to use ScalarOutcome objects for optimization.

Search over levers

Directed search is most often used to search over the decision levers in order to find good candidate strategies. This is for example the first step in the Many Objective Robust Decision Making process. This is straightforward to do with the workbench using the optimize method.

from ema_workbench import MultiprocessingEvaluator, ema_logging

ema_logging.log_to_stderr(ema_logging.INFO)

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.optimize(nfe=10000, searchover='levers', 
                                 epsilons=[0.1,]*len(model.outcomes),
                                 population_size=50)

the result from optimize is a DataFrame with the decision variables and outcomes of interest. The latest version of the workbench comes with a pure python implementation of parallel coordinates plot built on top of matplotlib. It has been designed with the matplotlib and seaborn api in mind. We can use this to quickly visualize the optimization results.

from ema_workbench.analysis import parcoords

paraxes = parcoords.ParallelAxes(parcoords.get_limits(results), rot=0)
paraxes.plot(results, color=sns.color_palette()[0])
paraxes.invert_axis('max_P')
plt.show()

Note how we can flip an axis using the invert_axis method. This eases interpretation of the figure because the ideal solution in this case would be a straight line for the four outcomes of interest at the top of the figure.

output_8_1

Specifying constraints

In the previous example, we showed the most basic way for using the workbench to perform many-objective optimization. However, the workbench also offers support for constraints and tracking convergence. Constrains are an attribute of the optimization problem, rather than an attribute of the model as in Rhodium. Thus, we can pass a list of constraints to the optimize method. A constraint can be applied to the model input parameters (both uncertainties and levers), and/or outcomes. A constraint is essentially a function that should return the distance from the feasibility threshold. The distance should be 0 if the constraint is met.

As a quick demonstration, let’s add a constraint on the maximum pollution. This constraint applies to the max_P outcome. The example below specifies that the maximum pollution should be below 1.

from ema_workbench import MultiprocessingEvaluator, ema_logging, Constraint

ema_logging.log_to_stderr(ema_logging.INFO)

constraints = [Constraint("max pollution", outcome_names="max_P",
                          function=lambda x:max(0, x-1))]

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.optimize(nfe=1000, searchover='levers', 
                                 epsilons=[0.1,]*len(model.outcomes),
                                 population_size=25, constraints=constraints)

tracking convergence

To track convergence, we need to specify which metric(s) we want to use and pass these to the optimize method. At present the workbench comes with 3 options: Hyper volume, Epsilon progress, and a class that will write the archive at each iteration to a separate text file enabling later processing. If convergence metrics are specified, optimize will return both the results as well as the convergence information.

from ema_workbench import MultiprocessingEvaluator, ema_logging
from ema_workbench.em_framework.optimization import (HyperVolume,
                                                     EpsilonProgress, )
from ema_workbench.em_framework.outcomes import Constraint

ema_logging.log_to_stderr(ema_logging.INFO)

# because of the constraint on pollution, we can specify the 
# maximum easily
convergence_metrics = [HyperVolume(minimum=[0,0,0,0], maximum=[1,1,1,1]),
                       EpsilonProgress()]
constraints = [Constraint("max pollution", outcome_names="max_P",
                          function=lambda x:max(0, x-1))]

with MultiprocessingEvaluator(model) as evaluator:
    results_ref1, convergence1 = evaluator.optimize(nfe=25000, searchover='levers', 
                                    epsilons=[0.05,]*len(model.outcomes),
                                    convergence=convergence_metrics,
                                    constraints=constraints,
                                    population_size=100)

We can visualize the results using parcoords as before, while the convergence information is in a DataFrame making it also easy to plot.

fig, (ax1, ax2) = plt.subplots(ncols=2, sharex=True)
ax1.plot(convergence1.epsilon_progress)
ax1.set_xlabel('nr. of generations')
ax1.set_ylabel('$\epsilon$ progress')
ax2.plot(convergence1.hypervolume)
ax2.set_ylabel('hypervolume')
sns.despine()
plt.show()

output_16_0

Changing the reference scenario

Up till now, we have performed the optimization for an unspecified reference scenario. Since the lake model function takes default values for each of the deeply uncertain factors, these values have been implicitly assumed. It is however possible to explicitly pass a reference scenario that should be used instead. In this way, it is easy to apply the extended MORDM approach suggested by Watson and Kasprzyk (2017).

To see the effects of changing the reference scenario on the values for the decision levers found through the optimization, as well as ensuring a fair comparison with the previous results, we use the same convergence metrics and constraints from the previous optimization. Note that the constraints are in essence only a function, and don’t retain optimization specific state, we can simply reuse them. The convergence metrics, in contrast retain state and we thus need to re-instantiate them.

from ema_workbench import Scenario

reference = Scenario('reference', **dict(b=.43, q=3,mean=0.02, 
                                         stdev=0.004, delta=.94))
convergence_metrics = [HyperVolume(minimum=[0,0,0,0], maximum=[1,1,1,1]),
                       EpsilonProgress()]

with MultiprocessingEvaluator(model) as evaluator:
    results_ref2, convergence2 = evaluator.optimize(nfe=25000, searchover='levers', 
                                  epsilons=[0.05,]*len(model.outcomes),
                                  convergence=convergence_metrics,
                                  constraints=constraints,
                                  population_size=100, reference=reference)

comparing results for different reference scenarios

To demonstrate the parcoords plotting functionality in some more detail, let’s combine the results from the optimizations for the two different reference scenarios and visualize them in the same plot. To do this, we need to first figure out the limits across both optimizations. Moreover, to get a better sense of which part of the decision space is being used, let’s set the limits for the decision levers on the basis of their specified ranges instead of inferring the limits from the optimization results.

columns = [lever.name for lever in model.levers]
columns += [outcome.name for outcome in model.outcomes]
limits = {lever.name: (lever.lower_bound, lever.upper_bound) for lever in 
           model.levers}
limits = dict(**limits, **{outcome.name:(0,1) for outcome in model.outcomes})
limits = pd.DataFrame.from_dict(limits)
# we resort the limits in the order produced by the optimization
limits = limits[columns] 

paraxes = parcoords.ParallelAxes(limits, rot=0)
paraxes.plot(results_ref1, color=sns.color_palette()[0], label='ref1')
paraxes.plot(results_ref2, color=sns.color_palette()[1], label='ref2')
paraxes.legend()
paraxes.invert_axis('max_P')
plt.show()

output_22_0.png

Robust Search

The workbench also comes with support for many objective robust optimization. In this case, each candidate solution is evaluated over a set of scenarios, and the robustness of the performance over this set is calculated. This requires specifying 2 new pieces of information:
* the robustness metrics
* the scenarios over which to evaluate the candidate solutions

The robustness metrics are simply a collection of ScalarOutcome objects. For each one, we have to specify which model outcome(s) it uses, as well as the actual robustness function. For demonstrative purposes, let’s assume we are use a robustness function using descriptive statistics: we want to maximize the 10th percentile performance for reliability, inertia, and utility, while minimizing the 90th percentile performance for max_P.

We can specify our scenarios in various ways. The simplest would be to pass the number of scenarios to the robust_optimize method. In this case for each generation a new set of scenarios is used. This can create noise and instability in the optimization. A better option is to explicitly generate the scenarios first, and pass these to the method. In this way, the same set of scenarios is used for each generation.

If we want to specify a constraint, this can easily be done. Note however, that in case of robust optimization, the constrains will apply to the robustness metrics instead of the model outcomes. They can of course still apply to the decision variables as well.

import functools
from ema_workbench import Constraint, MultiprocessingEvaluator
from ema_workbench import Constraint, ema_logging
from ema_workbench.em_framework.optimization import (HyperVolume,
                                                     EpsilonProgress)
from ema_workbench.em_framework.samplers import sample_uncertainties

ema_logging.log_to_stderr(ema_logging.INFO)

percentile10 = functools.partial(np.percentile, q=10)
percentile90 = functools.partial(np.percentile, q=90)

MAXIMIZE = ScalarOutcome.MAXIMIZE
MINIMIZE = ScalarOutcome.MINIMIZE
robustnes_functions = [ScalarOutcome('90th percentile max_p', kind=MINIMIZE, 
                             variable_name='max_P', function=percentile90),
                       ScalarOutcome('10th percentile reliability', kind=MAXIMIZE, 
                             variable_name='reliability', function=percentile10),
                       ScalarOutcome('10th percentile inertia', kind=MAXIMIZE, 
                             variable_name='inertia', function=percentile10),
                       ScalarOutcome('10th percentile utility', kind=MAXIMIZE, 
                             variable_name='utility', function=percentile10)]

def constraint(x):
    return max(0, percentile90(x)-10)

constraints = [Constraint("max pollution", 
                          outcome_names=['90th percentile max_p'],
                          function=constraint)]
convergence_metrics = [HyperVolume(minimum=[0,0,0,0], maximum=[10,1,1,1]),
                       EpsilonProgress()]
n_scenarios = 10
scenarios = sample_uncertainties(model, n_scenarios)

nfe = 10000

with MultiprocessingEvaluator(model) as evaluator:
    robust_results, convergence = evaluator.robust_optimize(robustnes_functions, 
                            scenarios, nfe=nfe, constraints=constraints,
                            epsilons=[0.05,]*len(robustnes_functions),
                            convergence=convergence_metrics,)
fig, (ax1, ax2) = plt.subplots(ncols=2)
ax1.plot(convergence.epsilon_progress.values)
ax1.set_xlabel('nr. of generations')
ax1.set_ylabel('$\epsilon$ progress')
ax2.plot(convergence.hypervolume)
ax2.set_ylabel('hypervolume')
sns.despine()
plt.show()

output_25_0.png

paraxes = parcoords.ParallelAxes(parcoords.get_limits(robust_results), rot=45)
paraxes.plot(robust_results)
paraxes.invert_axis('90th percentile max_p')
plt.show()

output_26_0.png

Search over uncertainties: worst case discovery

Up till now, we have focused on optimizing over the decision levers. The workbench however can also be used for worst case discovery (Halim et al, 2016). In essence, the only change is to specify that we want to search over uncertainties instead of over levers. Constraints and convergence works just as in the previous examples.

Reusing the foregoing, however, we should change the direction of optimization of the outcomes. We are no longer interested in finding the best possible outcomes, but instead we want to find the worst possible outcomes.

# change outcomes so direction is undesirable
minimize = ScalarOutcome.MINIMIZE
maximize = ScalarOutcome.MAXIMIZE

for outcome in model.outcomes:
    if outcome.kind == minimize:
        outcome.kind = maximize
    else:
        outcome.kind = minimize

We can reuse the reference keyword argument to perform worst case discovery for one of the policies found before. So, below we select solution number 9 from the pareto approximate set. We can turn this into a dict and instantiate a Policy objecti.

from ema_workbench import Policy

policy = Policy('9', **{k:v for k, v in results_ref1.loc[9].items()
                        if k in model.levers})

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.optimize(nfe=1000, searchover='uncertainties', 
                                 epsilons=[0.1,]*len(model.outcomes),
                                 reference=policy)

Visualizing the results is straightforward using parcoords.

paraxes = parcoords.ParallelAxes(parcoords.get_limits(results), rot=0)
paraxes.plot(results)
paraxes.invert_axis('max_P')
plt.show()

output_30_0.png

Closing remarks

This blog showcased the functionality of the workbench for applying search based approaches to exploratory modelling. We specifically looked at the use of many-objective optimization for searching over the levers or uncertainties, as well as the use of many-objective robust optimization. This completes the overview of the functionality available in the workbench. In the next blog, I will put it all together to show how the workbench can be used to perform Many Objective Robust Decision Making.

Advertisements

Open exploration with the Exploratory Modelling Workbench

In this blog, I will continue to showcase the functionality of the exploratory modelling workbench. In the previous blog, I have given a general introduction to the workbench, and showed how the Direct Policy Search example that comes with Rhodium can be adapted for use with the workbench. In this blog post, I will showcase how the workbench can be used for open exploration.

first a short background

In exploratory modeling, we are interested in understanding how regions in the uncertainty space and/or the decision space map to the whole outcome space, or partitions thereof. There are two general approaches for investigating this mapping. The first one is through systematic sampling of the uncertainty or decision space. This is sometimes also known as open exploration. The second one is to search through the space in a directed manner using some type of optimization approach. This is sometimes also known as directed search.

The workbench support both open exploration and directed search. Both can be applied to investigate the mapping of the uncertainty space and/or the decision space to the outcome space. In most applications, search is used for finding promising mappings from the decision space to the outcome space, while exploration is used to stress test these mappings under a whole range of possible resolutions to the various uncertainties. This need not be the case however. Optimization can be used to discover the worst possible scenario, while sampling can be used to get insight into the sensitivity of outcomes to the various decision levers.

open exploration

To showcase the open exploration functionality, let’s start with a basic example using the DPS lake problem also used in the previous blog post. We are going to simultaneously sample over uncertainties and decision levers. We are going to generate 1000 scenarios and 5 policies, and see how they jointly affect the outcomes. A scenario is understood as a point in the uncertainty space, while a policy is a point in the decision space. The combination of a scenario and a policy is called experiment. The uncertainty space is spanned by uncertainties, while the decision space is spanned by levers. Both uncertainties and levers are instances of RealParameter (a continuous range), IntegerParameter (a range of integers), or CategoricalParameter (an unorder set of things). By default, the workbench will use Latin Hypercube sampling for generating both the scenarios and the policies. Each policy will be always evaluated over all scenarios (i.e. a full factorial over scenarios and policies).

from ema_workbench import (RealParameter, ScalarOutcome, Constant,
                           ReplicatorModel)
model = ReplicatorModel('lakeproblem', function=lake_model)
model.replications = 150

#specify uncertainties
model.uncertainties = [RealParameter('b', 0.1, 0.45),
                       RealParameter('q', 2.0, 4.5),
                       RealParameter('mean', 0.01, 0.05),
                       RealParameter('stdev', 0.001, 0.005),
                       RealParameter('delta', 0.93, 0.99)]

# set levers
model.levers = [RealParameter("c1", -2, 2),
                RealParameter("c2", -2, 2),
                RealParameter("r1", 0, 2),
                RealParameter("r2", 0, 2),
                RealParameter("w1", 0, 1)]

def process_p(values):
    values = np.asarray(values)
    values = np.mean(values, axis=0)
    return np.max(values)

#specify outcomes
model.outcomes = [ScalarOutcome('max_P', kind=ScalarOutcome.MINIMIZE,
                                function=process_p),
                  ScalarOutcome('utility', kind=ScalarOutcome.MAXIMIZE,
                                function=np.mean),
                  ScalarOutcome('inertia', kind=ScalarOutcome.MINIMIZE,
                                function=np.mean),
                  ScalarOutcome('reliability', kind=ScalarOutcome.MAXIMIZE,
                                function=np.mean)]

# override some of the defaults of the model
model.constants = [Constant('alpha', 0.41),
                   Constant('steps', 100)]

Next, we can perform experiments with this model.

from ema_workbench import (MultiprocessingEvaluator, ema_logging,
                           perform_experiments)
ema_logging.log_to_stderr(ema_logging.INFO)

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.perform_experiments(scenarios=1000, policies=5)

Visual analysis

Having generated these results, the next step is to analyze them and see what we can learn from the results. The workbench comes with a variety of techniques for this analysis. A simple first step is to make a few quick visualizations of the results. The workbench has convenience functions for this, but it also possible to create your own visualizations using the scientific Python stack.

from ema_workbench.analysis import pairs_plotting
fig, axes = pairs_plotting.pairs_scatter(results, group_by='policy',
                                         legend=False)
plt.show()

output_6_0

Writing your own visualizations requires a more in-depth understanding of how the results from the workbench are structured. perform_experiments returns a tuple. The first item is a numpy structured array where each row is a single experiment. The second item contains the outcomes, structured in a dict with the name of the outcome as key and a numpy array as value. Experiments and outcomes are aligned based on index.

import seaborn as sns

experiments, outcomes = results

df = pd.DataFrame.from_dict(outcomes)
df = df.assign(policy=experiments['policy'])

# rename the policies using numbers
df['policy'] = df['policy'].map({p:i for i, p in
                                enumerate(set(experiments['policy']))})

# use seaborn to plot the dataframe
grid = sns.pairplot(df, hue='policy', vars=outcomes.keys())
ax = plt.gca()
plt.show()

output_8_0

Often, it is convenient to separate the process of performing the experiments from the analysis. To make this possible, the workbench offers convenience functions for storing results to disc and loading them from disc. The workbench will store the results in a tarbal with .csv files and separate metadata files. This is a convenient format that has proven sufficient over the years.

from ema_workbench import save_results

save_results(results, '1000 scenarios 5 policies.tar.gz')

from ema_workbench import load_results

results = load_results('1000 scenarios 5 policies.tar.gz')

advanced analysis

In addition to visual analysis, the workbench comes with a variety of techniques to perform a more in-depth analysis of the results. In addition, other analyses can simply be performed by utilizing the scientific python stack. The workbench comes with

  • Scenario Discovery, a model driven approach to scenario development.
  • Dimensional stacking, a quick visual approach drawing on feature scoring to enable scenario discovery. This approach has received limited attention in the literature (Suzuki et al., 2015). The implementation in the workbench replaces the rule mining approach with a feature scoring approach.
  • Feature Scoring, a poor man’s alternative to global sensitivity analysis
  • Regional sensitivity analysis

Scenario Discovery

A detailed discussion on scenario discovery can be found in an earlier blogpost. For completeness, I provide a code snippet here. Compared to the previous blog post, there is one small change. The library mpld3 is currently not being maintained and broken on Python 3.5 and higher. To still utilize the interactive exploration of the trade offs within the notebook, use the interactive back-end as shown below.

from ema_workbench.analysis import prim

experiments, outcomes = results

x = experiments
y = outcomes['max_P'] <0.8

prim_alg = prim.Prim(x, y, threshold=0.8)
box1 = prim_alg.find_box()

%matplotlib notebook

box1.show_tradeoff()
plt.show()

tradeoff

%matplotlib inline
# we go back to default not interactive

box1.inspect(43)
box1.inspect(43, style='graph')
plt.show()

output_13_1

dimensional stacking

Dimensional stacking was suggested as a more visual approach to scenario discovery. It involves two steps: identifying the most important uncertainties that affect system behavior, and creating a pivot table using the most influential uncertainties. Creating the pivot table involves binning the uncertainties. More details can be found in Suzuki et al. (2015) or by looking through the code in the workbench. Compared to the original paper, I use feature scoring for determining the most influential uncertainties. The code is set up in a modular way so other approaches to global sensitivity analysis can easily be used as well if so desired.

from ema_workbench.analysis import dimensional_stacking

x = experiments
y = outcomes['max_P'] <0.8

dimensional_stacking.create_pivot_plot(x,y, 2, nbins=3)
plt.show()

output_15_1

We can see from this visual that if B is low, while Q is high, we have a high concentration of cases where pollution stays below 0.8. The mean and delta have some limited additional influence. By playing around with an alternative number of bins, or different number of layers, patterns can be coarsened or refined.

regional sensitivity analysis

A third approach for supporting scenario discovery is to perform a regional sensitivity analysis. The workbench implements a visual approach based on plotting the empirical CDF given a classification vector. Please look at section 3.4 in Pianosi et al (2016) for more details.

from ema_workbench.analysis import regional_sa
from numpy.lib import recfunctions as rf

x = rf.drop_fields(experiments, 'model', asrecarray=True)
y = outcomes['max_P'] < 0.8

regional_sa.plot_cdfs(x,y)
plt.show()

output_17_0

feature scoring

Feature scoring is a family of techniques often used in machine learning to identify the most relevant features to include in a model. This is similar to one of the use cases for global sensitivity analysis, namely factor prioritisation. In some of the work ongoing in Delft, we are comparing feature scoring with Sobol and Morris and the results are quite positive. The main advantage of feature scoring techniques is that they impose virtually no constraints on the experimental design, while they can handle real valued, integer valued, and categorical valued parameters. The workbench supports multiple techniques, the most useful of which generally is extra trees (Geurts et al. 2006).

For this example, we run feature scoring for each outcome of interest. We can also run it for a specific outcome if desired. Similarly, we can choose if we want to run in regression mode or classification mode. The later is applicable if the outcome is a categorical variable and the results should be interpreted similar to regional sensitivity analysis results. For more details, see the documentation.

from ema_workbench.analysis import feature_scoring

x = experiments
y = outcomes

fs = feature_scoring.get_feature_scores_all(x, y)
sns.heatmap(fs, cmap='viridis', annot=True)
plt.show()

output_19_0

From the results, we see that max_P is primarily influenced by b, while utility is driven by delta, for inertia and reliability the situation is a little bit less clear cut.

linear regression

In addition to the prepackaged analyses that come with the workbench, it is also easy to rig up something quickly using the ever expanding scientific Python stack. Below is a quick example of performing a basic regression analysis on the results.

experiments, outcomes = results

for key, value in outcomes.items():
    params = model.uncertainties #+ model.levers[:]

    fig, axes = plt.subplots(ncols=len(params), sharey=True)

    y = value

    for i, param in enumerate(params):
        ax = axes[i]
        ax.set_xlabel(param.name)

        pearson = sp.stats.pearsonr(experiments[param.name], y)

        ax.annotate("r: {:6.3f}".format(pearson[0]), xy=(0.15, 0.85),
                    xycoords='axes fraction',fontsize=13)

        x = experiments[param.name]
        sns.regplot(x, y, ax=ax, ci=None, color='k',
        scatter_kws={'alpha':0.2, 's':8, 'color':'gray'})

        ax.set_xlim(param.lower_bound, param.upper_bound)

    axes[0].set_ylabel(key)

plt.show()

output_22_0

More advanced sampling techniques

The workbench can also be used for more advanced sampling techniques. To achieve this, it relies on SALib. On the workbench side, the only change is to specify the sampler we want to use. Next, we can use SALib directly to perform the analysis. To help with this, the workbench provides a convenience function for generating the problem dict which SALib provides. The example below focusses on performing SOBOL on the uncertainties, but we could do the exact same thing with the levers instead. The only changes required would be to set lever_sampling instead of uncertainty_sampling, and get the SALib problem dict based on the levers.

from SALib.analyze import sobol
from ema_workbench.em_framework.salib_samplers import get_SALib_problem

with MultiprocessingEvaluator(model) as evaluator:
    sa_results = evaluator.perform_experiments(scenarios=1000,
                                               uncertainty_sampling='sobol')

experiments, outcomes = sa_results
problem = get_SALib_problem(model.uncertainties)

Si = sobol.analyze(problem, outcomes['max_P'],
                   calc_second_order=True, print_to_console=False)

Si_filter = {k:Si[k] for k in ['ST','ST_conf','S1','S1_conf']}
Si_df = pd.DataFrame(Si_filter, index=problem['names'])

Using the Exploratory Modelling Workbench

Over the last 7 years, I have been working on the development of an open source toolkit for supporting decision-making under deep uncertainty. This toolkit is known as the exploratory modeling workbench. The motivation for this name is that in my opinion all model-based deep uncertainty approaches are forms of exploratory modeling as first introduced by Bankes (1993). The design of the workbench has undergone various changes over time, but it has started to stabilize in the fall of 2016. This summer, I published a paper detailing the workbench (Kwakkel, 2017). There is an in depth example in the paper, but in a series of blogs I want to showcase the funtionality in some more detail.

The workbench is readily available through pip, but it requires ipyparallel and mpld3 (both available through conda), SALib (via pip), and optionality platypus (pip install directly from github repo).

Adapting the DPS example from Rhodium

As a starting point, I will use the Direct Policy Search example that is available for Rhodium (Quinn et al 2017). I will adapt this code to work with the workbench. In this way, I can explain the workbench, as well as highlight some of the main differences between the workbench and Rhodium.

<br /># A function for evaluating our cubic DPS. This is based on equation (12)
# from [1].
def evaluateCubicDPS(policy, current_value):
    value = 0

for i in range(policy["length"]):
    rbf = policy["rbfs"][i]
    value += rbf["weight"] * abs((current_value - rbf["center"]) / rbf["radius"])**3
    value = min(max(value, 0.01), 0.1)
    return value

# Construct the lake problem
def lake_problem(policy, # the DPS policy
                 b = 0.42, # decay rate for P in lake (0.42 = irreversible)
                 q = 2.0, # recycling exponent
                 mean = 0.02, # mean of natural inflows
                 stdev = 0.001, # standard deviation of natural inflows
                 alpha = 0.4, # utility from pollution
                 delta = 0.98, # future utility discount rate
                 nsamples = 100, # monte carlo sampling of natural inflows
                 steps = 100): # the number of time steps (e.g., days)
    Pcrit = root(lambda x: x**q/(1+x**q) - b*x, 0.01, 1.5)
    X = np.zeros((steps,))
    decisions = np.zeros((steps,))
    average_daily_P = np.zeros((steps,))
    reliability = 0.0
    utility = 0.0
    inertia = 0.0

    for _ in range(nsamples):
        X[0] = 0.0

        natural_inflows = np.random.lognormal(
                math.log(mean**2 / math.sqrt(stdev**2 + mean**2)),
                math.sqrt(math.log(1.0 + stdev**2 / mean**2)),
                size=steps)

        for t in range(1,steps):
            decisions[t-1] = evaluateCubicDPS(policy, X[t-1])
            X[t] = (1-b)*X[t-1] + X[t-1]**q/(1+X[t-1]**q) + decisions[t-1] + natural_inflows[t-1]
            average_daily_P[t] += X[t]/float(nsamples)

        reliability += np.sum(X < Pcrit)/float(steps) 
        utility += np.sum(alpha*decisions*np.power(delta,np.arange(steps)))
        inertia += np.sum(np.diff(decisions) > -0.01)/float(steps-1)

    max_P = np.max(average_daily_P)
    reliability /= float(nsamples)
    utility /= float(nsamples)
    inertia /= float(nsamples)

    return (max_P, utility, inertia, reliability)

The formulation of the decision rule assumes that policy is a dict, which is composed of a set of variables generated either through sampling or through optimization. This is relatively straightforward to do in Rhodium, but not so easy to do in the workbench. In the workbench, a policy is a composition of policy levers, where each policy lever is either a range of real values, a range of integers, or an unordered set of categories. To adapt the DPS version of the lake problem to work with the workbench, we have to first replace the policy dict with the different variables explicitly.

def get_antropogenic_release(xt, c1, c2, r1, r2, w1):
    '''
    Parameters
    ----------
    xt : float
    polution in lake at time t
    c1 : float
    center rbf 1
    c2 : float
    center rbf 2
    r1 : float
    radius rbf 1
    r2 : float
    radius rbf 2
    w1 : float
    weight of rbf 1

    note:: w2 = 1 - w1

    '''

    rule = w1*(abs(xt-c1/r1))**3+(1-w1)*(abs(xt-c2/r2))**3
    at = min(max(rule, 0.01), 0.1)
    return at

Next, we need to adapt the lake_problem function itself to use this adapted version of the decision rule. This requires 2 changes: replace policy in the function signature of the lake_model function with the actual underlying parameters c1, c2, r1, r2, and w1, and use these when calculating the anthropological pollution rate.

def lake_model(b=0.42, q=2.0, mean=0.02, stdev=0.001, alpha=0.4, delta=0.98,
               c1=0.25, c2=0.25, r1=0.5, r2=0.5, w1=0.5, nsamples=100,
               steps=100):
    Pcrit = root(lambda x: x**q/(1+x**q) - b*x, 0.01, 1.5)
    X = np.zeros((steps,))
    decisions = np.zeros((steps,))
    average_daily_P = np.zeros((steps,))
    reliability = 0.0
    utility = 0.0
    inertia = 0.0

    for _ in range(nsamples):
        X[0] = 0.0

        natural_inflows = np.random.lognormal(
                math.log(mean**2 / math.sqrt(stdev**2 + mean**2)),
                math.sqrt(math.log(1.0 + stdev**2 / mean**2)),
                          size=steps)

        for t in range(1,steps):
            decisions[t-1] = get_antropogenic_release(X[t-1], c1, c2, r1, r2, w1)
            X[t] = (1-b)*X[t-1] + X[t-1]**q/(1+X[t-1]**q) + decisions[t-1] + natural_inflows[t-1]
            average_daily_P[t] += X[t]/float(nsamples)

        reliability += np.sum(X < Pcrit)/float(steps)
        utility += np.sum(alpha*decisions*np.power(delta,np.arange(steps)))
        inertia += np.sum(np.diff(decisions) > -0.01)/float(steps-1)

    max_P = np.max(average_daily_P)
    reliability /= float(nsamples)
    utility /= float(nsamples)
    inertia /= float(nsamples)

    return (max_P, utility, inertia, reliability)

This version of the code can be combined with the workbench already. However, we can clean it up a bit more if we want to. Note how there are 2 for loops in the lake model. The outer loop generates stochastic realizations of the natural inflow, while the inner loop calculates the the dynamics of the system given a stochastic realization. The workbench can be made responsible for this outer loop.

A quick note on terminology is in order here. I have a background in transport modeling. Here we often use discrete event simulation models. These are intrinsically stochastic models. It is standard practice to run these models several times and take descriptive statistics over the set of runs. In discrete event simulation, and also in the context of agent based modeling, this is known as running replications. The workbench adopts this terminology and draws a sharp distinction between designing experiments over a set of deeply uncertain factors, and performing replications of each experiment to cope with stochastic uncertainty.

Some other notes on the code:
* To aid in debugging functions, it is good practice to make a function deterministic. In this case we can quite easily achieve this by including an optional argument for setting the seed of the random number generation.
* I have slightly changed the formulation of inertia, which is closer to the mathematical formulation used in the various papers.
* I have changes the for loop over t to get rid of virtually all the t-1 formulations

 

from __future__ import division # python2
import math
import numpy as np
from scipy.optimize import brentq

def lake_model(b=0.42, q=2.0, mean=0.02, stdev=0.001, alpha=0.4,
               delta=0.98, c1=0.25, c2=0.25, r1=0.5, r2=0.5,
               w1=0.5, nsamples=100, steps=100, seed=None):
    '''runs the lake model for 1 stochastic realisation using specified
       random seed.

    Parameters
    ----------
    b : float
    decay rate for P in lake (0.42 = irreversible)
    q : float
    recycling exponent
    mean : float
    mean of natural inflows
    stdev : float
    standard deviation of natural inflows
    alpha : float
    utility from pollution
    delta : float
    future utility discount rate
    c1 : float
    c2 : float
    r1 : float
    r2 : float
    w1 : float
    steps : int
    the number of time steps (e.g., days)
    seed : int, optional
    seed for the random number generator
    '''
    np.random.seed(seed)

    Pcrit = brentq(lambda x: x**q/(1+x**q) - b*x, 0.01, 1.5)
    X = np.zeros((steps,))
    decisions = np.zeros((steps,))

    X[0] = 0.0

    natural_inflows = np.random.lognormal(
                math.log(mean**2 / math.sqrt(stdev**2 + mean**2)),
                math.sqrt(math.log(1.0 + stdev**2 / mean**2)),
                size=steps)

    for t in range(steps-1):
        decisions[t] = get_antropogenic_release(X[t], c1, c2, r1, r2, w1)
        X[t+1] = (1-b)*X[t] + X[t]**q/(1+X[t]**q) + decisions[t] + natural_inflows[t]

    reliability = np.sum(X < Pcrit)/steps
    utility = np.sum(alpha*decisions*np.power(delta,np.arange(steps)))

    # note that I have slightly changed this formulation to retain
    # consistency with the equations in the papers
    inertia = np.sum(np.abs(np.diff(decisions)) < 0.01)/(steps-1)
    return X, utility, inertia, reliability

Now we are ready to connect this model to the workbench. This is fairly similar to how you would do it with Rhodium. We have to specify the uncertainties, the outcomes, and the policy levers. For the uncertainties and the levers, we can use real valued parameters, integer valued parameters, and categorical parameters. For outcomes, we can use either scalar, single valued outcomes or time series outcomes. For convenience, we can also explicitly control constants in case we want to have them set to a value different from their default value.

In this particular case, we are running the replications with the workbench. We still have to specify the descriptive statistics we would like to gather over the set of replications. For this, we can pass a function to an outcome. This function will be called with the results over the set of replications.

import numpy as np
from ema_workbench import (RealParameter, ScalarOutcome, Constant,
                           ReplicatorModel)

model = ReplicatorModel('lakeproblem', function=lake_model)
model.replications = 150

#specify uncertainties
model.uncertainties = [RealParameter('b', 0.1, 0.45),
                       RealParameter('q', 2.0, 4.5),
                       RealParameter('mean', 0.01, 0.05),
                       RealParameter('stdev', 0.001, 0.005),
                       RealParameter('delta', 0.93, 0.99)]

# set levers
model.levers = [RealParameter("c1", -2, 2),
                RealParameter("c2", -2, 2),
                RealParameter("r1", 0, 2),
                RealParameter("r2", 0, 2),
                RealParameter("w1", 0, 1)]

def process_p(values):
    values = np.asarray(values)
    values = np.mean(values, axis=0)
    return np.max(values)

#specify outcomes
model.outcomes = [ScalarOutcome('max_P', kind=ScalarOutcome.MINIMIZE,
                                function=process_p),
                  ScalarOutcome('utility', kind=ScalarOutcome.MAXIMIZE,
                                function=np.mean),
                  ScalarOutcome('inertia', kind=ScalarOutcome.MINIMIZE,
                                function=np.mean),
                  ScalarOutcome('reliability', kind=ScalarOutcome.MAXIMIZE,
                                function=np.mean)]

# override some of the defaults of the model
model.constants = [Constant('alpha', 0.41),
                   Constant('steps', 100)]

Open exploration

Now that we have specified the model with the workbench, we are ready to perform experiments on it. We can use evaluators to distribute these experiments either over multiple cores on a single machine, or over a cluster using ipyparallel. Using any parallelization is an advanced topic, in particular if you are on a windows machine. The code as presented here will run fine in parallel on a mac or Linux machine. If you are trying to run this in parallel using multiprocessing on a windows machine, from within a jupyter notebook, it won’t work. The solution is to move the lake_model and get_antropogenic_release to a separate python module and import the lake model function into the notebook.

Another common practice when working with the exploratory modeling workbench is to turn on the logging functionality that it provides. This will report on the progress of the experiments, as well as provide more insight into what is happening in particular in case of errors.

If we want to perform experiments on the model we have just defined, we can use the perform_experiments method on the evaluator, or the stand alone perform_experiments function. We can perform experiments over the uncertainties and/or over the levers. Any policy is evaluated over each of the scenarios. So if we want to use 100 scenarios and 10 policies, this means that we will end up performing 100 * 10 = 1000 experiments. By default, the workbench uses Latin hypercube sampling for both sampling over levers and sampling over uncertainties. However, the workbench also offers support for full factorial, partial factorial, and Monte Carlo sampling, as well as wrappers for the various sampling schemes provided by SALib.

from ema_workbench import (MultiprocessingEvaluator, ema_logging,
                           perform_experiments)
ema_logging.log_to_stderr(ema_logging.INFO)

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.perform_experiments(scenarios=10, policies=10)

Directed Search

Similarly, we can easily use the workbench to search for a good candidate strategy. This requires that platypus is installed. If platypus is installed, we can simply use the optimize method. By default, the workbench will use $\epsilon$-NSGAII. The workbench can be used to search over the levers in order to find a good candidate strategy as is common in Many-Objective Robust Decision Making. The workbench can also be used to search over the uncertainties in order to find for example the worst possible outcomes and the conditions under which they appear. This is a form of worst case discovery. The optimize method takes an optional reference argument. This can be used to set the scenario for which you want to find good policies, or for setting the policy for which you want to find the worst possible outcomes. This makes implementing the approach suggested in Watson & Kasprzyk (2017) very easy.

with MultiprocessingEvaluator(model) as evaluator:
    results = evaluator.optimize(nfe=1000, searchover='levers',
                                 epsilons=[0.1,]*len(model.outcomes))

Robust optimization

A third possibility is to perform robust optimization. In this case, the search will take place over the levers, but a given policy is than evaluated for a set of scenarios and the performance is defined over this set. To do this, we need to explicitly define robustness. For this, we can use the outcome object we have used before. In the example below we are defining robustness as the worst 10th percentile over the set of scenarios. We need to pass a variable_name argument to explicitly link outcomes of the model to the robustness metrics.

import functools

percentile10 = functools.partial(np.percentile, q=10)
percentile90 = functools.partial(np.percentile, q=90)

MAXIMIZE = ScalarOutcome.MAXIMIZE
MINIMIZE = ScalarOutcome.MINIMIZE
robustnes_functions = [ScalarOutcome('90th percentile max_p', kind=MINIMIZE,
                                     variable_name='max_P', function=percentile90),
                       ScalarOutcome('10th percentile reliability', kind=MAXIMIZE,
                                     variable_name='reliability', function=percentile10),
                       ScalarOutcome('10th percentile inertia', kind=MAXIMIZE,
                                     variable_name='inertia', function=percentile10),
                       ScalarOutcome('10th percentile utility', kind=MAXIMIZE,
                                     variable_name='utility', function=percentile10)]

Given the specification of the robustness function, the remainder is straightforward and analogous to normal optimization.

<br />n_scenarios = 200
scenarios = sample_uncertainties(lake_model, n_scenarios)
nfe = 100000

with MultiprocessingEvaluator(lake_model) as evaluator:
    robust_results = evaluator.robust_optimize(robustnes_functions, scenarios,
                                               nfe=nfe, epsilons=[0.05,]*len(robustnes_functions))

This blog has introduced the exploratory modeling workbench and has shown its basic functionality for sampling or searching over uncertainties and levers. In subsequent blogs, I will take a more in depth look at this funcitonality, as well as demonstrate how the workbench facilitates the entire Many-Objective Robust Decision Making process.

Scenario discovery in Python

The purpose of this blog post is to demonstrate how one can do scenario discovery in python. This blogpost will use the exploratory modeling workbench available on github. I will demonstrate how we can perform both PRIM in an interactive way, as well as briefly show how to use CART, which is also available in the exploratory modeling workbench. There is ample literature on both CART and PRIM and their relative merits for use in scenario discovery. So I won’t be discussing that here in any detail. This blog was first written as an ipython notebook, which can be found here

The workbench is mend as a one stop shop for doing exploratory modeling, scenario discovery, and (multi-objective) robust decision making. To support this, the workbench is split into several packages. The most important packages are expWorkbench that contains the support for setting up and executing computational experiments or (multi-objective) optimization with models; The connectors package, which contains connectors to vensim (system dynamics modeling package), netlogo (agent based modeling environment), and excel; and the analysis package that contains a wide range of techniques for visualization and analysis of the results from series of computational experiments. Here, we will focus on the analysis package. It some future blog post, I plan to demonstrate the use of the workbench for performing computational experimentation and multi-objective (robust) optimization.

The workbench can be found on github and downloaded from there. At present, the workbench is only available for python 2.7. There is a seperate branch where I am working on making a version of the workbench that works under both python 2.7 and 3. The workbench is depended on various scientific python libraries. If you have a standard scientific python distribution, like anaconda, installed, the main dependencies will be met. In addition to the standard scientific python libraries, the workbench is also dependend on deap for genetic algorithms. There are also some optional dependencies. These include seaborn and mpld3 for nicer and interactive visualizations, and jpype for controlling models implemented in Java, like netlogo, from within the workbench.

In order to demonstrate the use of the exploratory modeling workbench for scenario discovery, I am using a published example. I am using the data used in the original article by Ben Bryant and Rob Lempert where they first introduced scenario discovery. Ben Bryant kindly made this data available for my use. The data comes as a csv file. We can import the data easily using pandas. columns 2 up to and including 10 contain the experimental design, while the classification is presented in column 15

import pandas as pd

data = pd.DataFrame.from_csv('./data/bryant et al 2010 data.csv',
                             index_col=False)
x = data.ix[:, 2:11]
y = data.ix[:, 15]

The exploratory modeling workbench is built on top of numpy rather than pandas. This is partly a path dependecy issue. The earliest version of prim in the workbench is from 2012, when pandas was still under heavy development. Another problem is that the pandas does not contain explicit information on the datatypes of the columns. The implementation of prim in the exploratory workbench is however datatype aware, in contrast to the scenario discovery toolkit in R. That is, it will handle categorical data differently than continuous data. Internally, prim uses a numpy structured array for x, and a numpy array for y. We can easily transform the pandas dataframe to either.

x = x.to_records()
y = y.values

the exploratory modeling workbench comes with a seperate analysis package. This analysis package contains prim. So let’s import prim. The workbench also has its own logging functionality. We can turn this on to get some more insight into prim while it is running.

from analysis import prim
from expWorkbench import ema_logging
ema_logging.log_to_stderr(ema_logging.INFO);

Next, we need to instantiate the prim algorithm. To mimic the original work of Ben Bryant and Rob Lempert, we set the peeling alpha to 0.1. The peeling alpha determines how much data is peeled off in each iteration of the algorithm. The lower the value, the less data is removed in each iteration. The minimium coverage threshold that a box should meet is set to 0.8. Next, we can use the instantiated algorithm and find a first box.

prim_alg = prim.Prim(x, y, threshold=0.8, peel_alpha=0.1)
box1 = prim_alg.find_box()

Let’s investigate this first box is some detail. A first thing to look at is the trade off between coverage and density. The box has a convenience function for this called show_tradeoff. To support working in the ipython notebook, this method returns a matplotlib figure with some additional information than can be used by mpld3.

import matplotlib.pyplot as plt

box1.show_tradeoff()
plt.show()

fig1

The notebook contains an mpld3 version of the same figure with interactive pop ups. Let’s look at point 21, just as in the original paper. For this, we can use the inspect method. By default this will display two tables, but we can also make a nice graph instead that contains the same information.

box1.inspect(21)
box1.inspect(21, style='graph')
plt.show()

This first displays two tables, followed by a figure

coverage    0.752809
density     0.770115
mass        0.098639
mean        0.770115
res dim     4.000000
Name: 21, dtype: float64

                            box 21
                               min         max     qp values
Demand elasticity        -0.422000   -0.202000  1.184930e-16
Biomass backstop price  150.049995  199.600006  3.515113e-11
Total biomass           450.000000  755.799988  4.716969e-06
Cellulosic cost          72.650002  133.699997  1.574133e-01

fig 2

If one where to do a detailed comparison with the results reported in the original article, one would see small numerical differences. These differences arise out of subtle differences in implementation. The most important difference is that the exploratory modeling workbench uses a custom objective function inside prim which is different from the one used in the scenario discovery toolkit. Other differences have to do with details about the hill climbing optimization that is used in prim, and in particular how ties are handled in selecting the next step. The differences between the two implementations are only numerical, and don’t affect the overarching conclusions drawn from the analysis.

Let’s select this 21 box, and get a more detailed view of what the box looks like. Following Bryant et al., we can use scatter plots for this.

box1.select(21)
fig = box1.show_pairs_scatter()
fig.set_size_inches((12,12))
plt.show()

fig3

We have now found a first box that explains close to 80% of the cases of interest. Let’s see if we can find a second box that explains the remainder of the cases.

box2 = prim_alg.find_box()

The logging will inform us in this case that no additional box can be found. The best coverage we can achieve is 0.35, which is well below the specified 0.8 threshold. Let’s look at the final overal results from interactively fitting PRIM to the data. For this, we can use to convenience functions that transform the stats and boxes to pandas data frames.

print prim_alg.stats_to_dataframe()
print prim_alg.boxes_to_dataframe()
       coverage   density      mass  res_dim
box 1  0.752809  0.770115  0.098639        4
box 2  0.247191  0.027673  0.901361        0
                             box 1              box 2
                               min         max    min         max
Demand elasticity        -0.422000   -0.202000   -0.8   -0.202000
Biomass backstop price  150.049995  199.600006   90.0  199.600006
Total biomass           450.000000  755.799988  450.0  997.799988
Cellulosic cost          72.650002  133.699997   67.0  133.699997

For comparison, we can also use CART for doing scenario discovery. This is readily supported by the exploratory modelling workbench.

from analysis import cart
cart_alg = cart.CART(x,y, 0.05)
cart_alg.build_tree()

Now that we have trained CART on the data, we can investigate its results. Just like PRIM, we can use stats_to_dataframe and boxes_to_dataframe to get an overview.

print cart_alg.stats_to_dataframe()
print cart_alg.boxes_to_dataframe()
       coverage   density      mass  res dim
box 1  0.011236  0.021739  0.052154        2
box 2  0.000000  0.000000  0.546485        2
box 3  0.000000  0.000000  0.103175        2
box 4  0.044944  0.090909  0.049887        2
box 5  0.224719  0.434783  0.052154        2
box 6  0.112360  0.227273  0.049887        3
box 7  0.000000  0.000000  0.051020        3
box 8  0.606742  0.642857  0.095238        2
                       box 1                  box 2               box 3  \
                         min         max        min         max     min
Cellulosic yield        80.0   81.649994  81.649994   99.900002  80.000
Demand elasticity       -0.8   -0.439000  -0.800000   -0.439000  -0.439
Biomass backstop price  90.0  199.600006  90.000000  199.600006  90.000   

                                         box 4                box 5  \
                               max         min         max      min
Cellulosic yield         99.900002   80.000000   99.900002   80.000
Demand elasticity        -0.316500   -0.439000   -0.316500   -0.439
Biomass backstop price  144.350006  144.350006  170.750000  170.750   

                                      box 6                  box 7  \
                               max      min         max        min
Cellulosic yield         99.900002  80.0000   89.050003  89.050003
Demand elasticity        -0.316500  -0.3165   -0.202000  -0.316500
Biomass backstop price  199.600006  90.0000  148.300003  90.000000   

                                         box 8
                               max         min         max
Cellulosic yield         99.900002   80.000000   99.900002
Demand elasticity        -0.202000   -0.316500   -0.202000
Biomass backstop price  148.300003  148.300003  199.600006

Alternatively, we might want to look at the classification tree directly. For this, we can use the show_tree method. This returns an image that we can either save, or display.

fig3

If we look at the results of CART and PRIM, we can see that in this case PRIM produces a better description of the data. The best box found by CART has a coverage and density of a little above 0.6. In contrast, PRIM produces a box with coverage and density above 0.75.