on 07-Jul-2022 (Thu)

Causal edges assumption is asymmetric; “ 𝑋 is a cause of 𝑌 ” is not the same as saying “ 𝑌 is a cause of 𝑋

For every agent-based model, it is necessary to define the agents and the environment of the modelled system via certain qualitative or quantitative properties and find rules or equations that govern how the agents interact with each other and with their environment. The rules should include what information the agents have access to and what action they then take considering bounded rationality [21–23]. While the input the agents use for their decision is relatively easy to find and to justify, specifying the rules that lead to a decision is much more difficult and often relies on assumptions from psychology or economics, which are often hard to back up empirically or by theories. This makes the search for valid rules for agent behaviour to one of the biggest challenges in agent-based modelling

Universal Framework for Agent based Models - 4 phases

(1) Initialization (2) Experience(3) Training(4) Application

Conda is a package, dependency, and environment management tool for Anaconda Python, which is widely used in the scientific community, especially on the Windows platform where the installation of binary extensions can be difficult.

Conda helps manage Python dependencies in two primary ways:

• Allows the creation of environments that isolate each project, thereby preventing dependency conflicts between projects.
• Provides identification of dependency conflicts at time of package installation, thereby preventing conflicts within projects/environments.

How Does Conda Compare to Pip, Virtualenv, Venv & Pyenv

Conda provides many of the features found in pip, virtualenv, venv and pyenv. However it is a completely separate tool that will manage Python dependencies differently, and only works in Conda environments.

Conda analyzes each package for compatible dependencies, and how to install them without conflict. If there is a conflict, Conda will let you know that the installation cannot be completed. By comparison, Pip installs all package dependencies regardless of whether they conflict with other packages already installed. To avoid dependency conflicts, use tools such as virtualenv, venv or pyenv to create isolated Anaconda environments.

For information about the use of pip in conda environments, refer to this Quickread post. How to Add Packages in Anaconda Python: Conda Vs. Pip.

Recommendations for Avoiding Dependency Conflicts with Conda

There are two simple rules to follow:

1. Always create a new environment for each project
2. Install all the packages that you need in the new environment at the same time. Installing packages one at a time can lead to dependency conflicts.

To create an environment with a specific version of Python and multiple packages including a package with a specific version:

$ conda create -n <env_name> python=<version#> <packagename> <packagename> <packagename>=<version#>

Alternatively, you can use conda to install all the packages in a requirements.txt file. You can save a requirements.txt file from an existing environment, or manually create a new requirements.txt for a different environment.

The common approach is to make use of relatively simple agent models (for example, based on qualitative knowledge of the domain, qualitative understanding of human behavior, etc.), so that complexity arises primarily from agent interactions among themselves and with the environment. For example, Thiele et al. [40] document that only 14% of articles published in the Journal of Artificial Societies and Social Sim- ulation include parameter fitting. Our key methodological contribution is a departure from developing simple agent models based on relevant qualitative insights to learning such models entirely on data. Due to its reliance on data about individual agent behavior, our approach is not universally applicable.

Our proposal of calibration at the agent level, in contrast, enables us to leverage state-of-the-art machine learning techniques, as well as obtain more reliable, and interpretable, models at the individual agent level. Recently, in field of ecology and sociology, there is rising interest to combine agent-based model with empirical methods [24]. Biophysical measurements, i.e., soil properties and socioeconomic surveys are used by Berger and Schreinemachers [4] to generate a landscape and agent populations which are consistent with empirical observation by Monte Carlo techniques. Notice that this is quite different application from ours, since we do not need to generate an agent population; rather we instantiate our multi-agent simulation with learned agents.

We offer instead a framework for data-driven agent-based modeling (DDABM), where agent models are learned from data about individual (typically, human) behavior, and the agent-based model is thereby fully data-driven, with no additional parameters to govern its behavior.

We now present our general framework for data-driven agent-based modeling (DDABM), which we subsequently apply to the problem of modeling residential rooftop solar diffusion in San Diego county, California. The key features of this framework are: a) explicit division of data into “calibration” and “validation” to ensure sound and reliable model validation and b) automated agent model training and cross-validation. In this framework, we make three assumptions. The first is that time is discrete. While this assumption is not of fundamental importance, it will help in presenting the concepts, and is the assumption made in our application. The second assumption is that agents are homogeneous. This may seem a strong assumption, but in fact it is without loss of generality. To see this, suppose that h(x) is our model of agent behaviour, where x is state, or all information that conditions the agent’s decision. Heterogeneity can be embedded in h by considering individual characteristics in state x, such as personality traits and socio-economic status, or, as in our application domain, housing characteristics.

Our third assumption is that each individual makes independent decisions at each time t, conditional on state x. Again, if x includes all features relevant to an agent’s decision, this assumption is relatively innocuous

Given these assumptions, DDABM proceeds as follows. We start with a data set of individual agent behavior over time, D = {(x it , y it )} i,t=0,...,T , where i indexes agents, t time through some horizon T and y it indicates agent i’s decision, i.e., 1 for “adopted” and 0 for “did not adopt” at time t. 1. Split the data D into calibration D c and validation D v parts along the time dimension: D c = {(x it , y it )} i,t≤T c and D v = {(x it , y it )} i,t>T c where T c is a time threshold. 2. Learn a model of agent behavior h on D c . Use cross-validation on D c for model (e.g., feature) selection. 3. Instantiate agents in the ABM using h learned in step 2. 4. Initialize the ABM to state x jT c for all artificial agents j. 5. Validate the ABM by running it from x T c using D v . One may wonder how to choose the initial state x jT c for the artificial agents. This is direct if the artificial agents in the ABM correspond to actual agents in the data. For example, in rooftop solar adoption we know which agents have adopted solar at time T c , and their actual housing characteristics, etc. Alternatively, one can run the ABM from the initial state, and start validation upon reaching time T c +

Developing Physically | i Cognitive Neuroscience The Biology of the Mind MICHAEL S. GAZZANIGA University of California, Santa Barbara RICHARD B.