This software is intended to help in designing and running simulations involving agents and their cognitive capacities.

An agent-based model (ABM) (also sometimes related to the term multi-agent system or multi-agent simulation) is a class of computational models for simulating the actions and interactions of autonomous agents (both individual or collective entities such as organizations or groups) with a view to assessing their effects on the system as a whole. It combines elements of game theory, complex systems, emergence, computational sociology, multi-agent systems, and evolutionary programming. Monte Carlo Methods are used to introduce randomness¹.

The main aim of this project proposal is to describe how perceptually grounded categories can emerge among the members of a population to allow successful communication and improve the cognitive performance of both the group as a whole, and in parallel each individual member of the group. The language plays a crucial role in this process by naming games that pinpoint the constructed categories and allow for effective sharing of those categories within the community of interacting individuals. Therefore, it influences both the dynamics of categorization process, and the final selection of categories that are stable in time, and shared within all members of a population.

Humans are equipped in senses that are the physiological capacities within organisms that provide inputs for perception. The biological structure of senses, the way how they process incoming input from external world, the classification of available stimulus repertoires, categorization, overlapping of similar ones, and corrections of natural noise were studied by neurobiology, neuroscience, cognitive psychology, and human physiology. In last decades several mathematical models of human senses were proposed. Those models can be implemented as the software algorithms and allow for the construction of artificial agents that can be further used for modeling of the role of human perception system in categorization processes. In addition several works of Steels proved the possibility to couple those models of human perception with the dynamics of naming process, therefore simulating in silico how a perceptually grounded categorical repertoire can become sufficiently shared among the members of a population². The population of artificial agents was able to build color categories and share them so that one agent from the population can use the word to get another agent to pick out the proper object from a set of colored objects in a cognitive scene³,⁴.

In our research proposal we address two issues: the dynamics of categorization process (i.e. equilibration process, where the time dependent changes of categorical repertoire in a group are observed), and the role of naming (i.e. as the natural interaction substrate shared between individuals in a population). Each senses perception and categorization of the input sensory information will be investigated, comparison between artificial agents models and human populations will be performed. The several types of internal Agents architecture will be tested, including advanced learning algorithms, such as RC Reservoir Computing (ESN Echo State Network, LSM Liquid State Machines⁵). Moreover the set of interacting Agents itself will be modeled using our in-house modification of RC algorithm that builds the ensemble of interacting LSM. So-called Liquid State Society (LSS) framework is able to model more efficiently the category formation and exchange, naming process in social learning systems. This model will be later used in mixed human-agent populations for better understanding the coupling between the categorization and the naming processes simulated using the senses models. Such experimental setups allow for selecting, which factors are crucial for successful communication events, and proper calibration of the mathematical models of considered senses.

Can be found on ReadTheDocs.

To run this software you need Python in version 2.x (3.x is not supported).

There are mostly libraries which helps you develop your own simulations. Sample simulation you might find in cog_simulations/steels/ folder. As of version 0.3 all parameters and configuration of simulations is stored in JSON files.

An example experiment. It was described in work of Steels, L. and Belpaeme, T. (2005). Sample configuration is in files examples/simulations/basic_simulation.json To run one go to cog_simulations folder and run: python steels/steels_main.py -p "../examples/simulations/basic_simulation.json" this should generate some files – some of them are used by presenter which visualizes result by showing agents categorization at given points, and one is used by analyzer

Usage: python steels_main.py [options]

Options:

-h, --help show this help message and exit -v, --verbose Increase verbosity (specify multiple times for more) -f FILE, --file=FILE output file with results -p PARAM_FILE, --params_file=PARAM_FILE file with parameters

Helps you analyze data generated during simulation. Normally it outputs data as table separated with tabulation, but with option -c you can change output to chart. Important thing: if you would like to see number of iteration when given statistic where taken you should specify to use statistic “it” (see examples)

Usage: python analyzer.py [-c] [-v] -f FILE statistic1 statistic2 ... where statistic in {cc;CSA;it;cv;CS;DS;DSA}

Options:

-h, --help show this help message and exit -v, --verbose increase verbosity (specify multiple times for more) -c, --chart specifies output to be a chart -f FILE, --file=FILE input file with results. THIS OPTION IS NECESSARY! --xlabel=XLABEL Label of x-axis --ylabel=YLABEL Label of y-axis

This program uses files with extension .pout . It presents agents categorization at given iteration. To run it go to cog_simulations/presenter Usage: munsell_palette.py [options]

Options:

-h, --help show this help message and exit -a AGENTS, --agents=AGENTS Number of agents viewed -d DIRECTORY, --directory=DIRECTORY Directory with input data -f FIND_FOCAL, --findfocal=FIND_FOCAL Determines which 'find_focal' algorithm will be used ('normal' as default or 'strength_based') -l LEGEND, --legend=LEGEND Type true to show language sharing

Run simulation. From /cog_simulations/steels run: python steels_main.py -p "../../examples/simulations/basic_simulation.json"

now see result in chart: python analyzer.py -f *.result -c it DS CS --xlabel="iteration" --ylabel "discriminative success & communicative success"

to see categorization: from /cog_simulations/presenter run: python munsell_palette.py -d ../ -a 10 --findfocal strength_based -l t

• numpy
• scipy
• progressbar
• matplotlib (for charts)
• grapefruit (built in?)
• pygraph (named python-graph-core)

Can be found here.

Authors

Programming: Konrad Kurdej, Michał Łukasik, Marek Maj, Przemysław Wróblewski, Julian Zubek

Mentoring: Dariusz Plewczyński, Franciszek Rakowski

This software is fork of COG-ABM from account cogcomp.

This software is licensed under LGPL-3.0 license.

https://www.gnu.org/licenses/lgpl-3.0.en.html