Input¶
Plans¶
Plans are the representation of agents’ possible actions. In EPOS, the plans are in form of n-dimensional vectors, which are read from the datasets. In each dataset, each agent has a certain number of plans, and each plan will have a certain of dimensions. A typical plan format is:
planScore : planDimension1, … , planDimensionN
in which the plan score can be:
- discomfort (Discomfort-based Local Cost Function) of the plan, or
- preference (Preference-based Local Cost Function) of the plan
The possible dataset formats are defined in agent.dataset package in the code. In addition, there is a Gaussian plan generator in agent.dataset.GaussianDataSet.java. For this dataset, you can define the number of agents, number of plans, and number of dimensions using the configuration or the Configuration Parameters.
In addition, at each iteration of EPOS, agents select their preliminary plans and report them to their parent. The implemented classes for this can be found at agent.planselection package. The main class is MultiObjectiveIeposPlanSelector implements PlanSelector<MultiObjectiveIEPOSAgent> and the main function is:
int id = agent.getOptimization().argmin(agent.getGlobalCostFunction(),
agent.getLocalCostFunction(),
agent.getPossiblePlans(),
otherResponse,
agent.getUnfairnessWeight(),
agent.getLocalCostWeight(),
currentDiscomfortSum,
currentDiscomfortSumSqr,
agent.getNumAgents(),
agent);
which return the id of the selected local plan. There are different ways of selecting the local plan, however by default EPOS chooses the plan with the PlanSelectionOptimizationFunctionCollection.complexFunction1. The function works as the following: \((1 - \alpha - \beta) \cdot global\_cost + \alpha \cdot unfairness + \beta \cdot local\_cost\), the details of which can be found in Multi-Objective Optimization.
Incentive Signals¶
EPOS is able to guide the selections of agents towards a certain predefined goal. The goal is given as a reference incentive signal, and the agents make selections towards the best matching between their joint selections and the reference signal. As EPOS is a minimization algorithm, the objective is the minimization of dissimilarity between them. For more information on the dissimilarity measures, please see Global Cost Function. The incentive signal is taken into account only if one of the following global cost functions are employed: Cross-Correlation, Residual Sum of Squares or Root Mean Square Error.
The Incentive signal is used within EPOS via the parameter goalSignalPath by setting a path to a file containing the incentive signal. If the file path is not provided, a first “.target” file in dataset directory will be used as a goal signal. If none “.target” files are found, a zero goal signal is constructed (a signal filled with zeros).
Note that the incentive signal must be of the same length as plans. If it is shorter than the planDim parameter, it will be padded at the end in a Round-Robyn fashion, if longer, it will be cropped. The signal is given in a column, one value in each row. The file cannot have any headers or more than one column.
For example, the following are correct ways of providing the incentive signal to EPOS:
goalSignalPath="/home/user/Documents/EPOS-project/datasets/incentive-signal.csv" // full path
goalSignalPath="datasets/incentive-signal.csv" // relative path
or in Configuration.java class:
public static String goalSignalFilename = "/home/user/Documents/EPOS-project/datasets/incentive-signal.csv"; // or
public static String goalSignalFilename = "datasets/incentive-signal.csv";
And these are the incorrect ways:
goalSignalPath="datasets" // the name of the directory containing the file is invalid
goalSignalPath="/home/user/Documents/EPOS-project/datasets/incentive-signal" // the file extension must be specified