De Jong, E.D. and L. Steels (2004). A Distributed Learning Algorithm
for Communication Development. Complex Systems, vol. 14, no. 4.
This appendix is also available as a single tar file: appendix.tar (100K).
Contents:
Running the algorithm
Output
Help information
File list
Sample output
References
Contact information
Running the algorithm
The algorithm can be run by starting Matlab(R) from the directory
containing this distribution and typing
cycle(nosteps, [, temp [, seed] ]);
where nosteps is the number of time steps for which the algorithm will
be run, the temperature parameter regulates the degree of exploration,
and seed is a parameter specifying the random seed (the
default is zero). The second and third parameters are optional.
For example, to run the algorithm for 1,000 cycles,
type:
cycle(1000);
To specify the temperature, use:
cycle(1000, 0.15);
And to additionally specify the random seed:
cycle(1000, 0.15, 17);
Output
As output, the algorithm displays the understanding matrix at the end
of the run and the values of the three measures: specificity,
consistency, and fidelity. The understanding matrix specifies, for
each pair of referents (ri, rj), the average probability that when
referent ri is encoded by an agent and decoded by another agent, the
result is referent rj. For successful communication, this matrix
should be close to the identity matrix. The fidelity measure is the
average of the diagonal of this matrix. Furthermore, a graph of the
average fidelity over time is plotted.
Online Matlab(R) help information
To get help on any of the functions, type Help function-name in
Matlab(R), e.g. Help cycle. Alternatively, one can inspect the file
directly (see list below). The algorithm corresponds directly to the
pseudocode described in the paper. The files contain commentary
so that, in combination with background information in the article,
variations of the algorithm may be explored. The main function is
cycle. It follows the description in the article, and contains
comments that explain its workings.
File list
boltz.m
compute_frequencies.m
consistency.m
cycle.m
determine_meaning_from_sensors.m
determine_meaning_from_words.m
fidelity.m
initialize_concepts.m
initialize_words.m
logdata.m
relprob.m
select.m
specificity.m
update_spec.m
update_use.m
word_production.m
Sample output
First, we run the algorithm once for 1,000 time steps. This is done with the command cycle(1000);. The output should be the following:
understanding =
| 0.9961 | 0.0013 | 0.0014 | 0.0013 |
| 0.0013 | 0.9962 | 0.0013 | 0.0013 |
| 0.0013 | 0.0013 | 0.9960 | 0.0014 |
| 0.0013 | 0.0013 | 0.0013 | 0.9962 |
Other measures can be obtained by monitoring the production matrices P(w|r) (avgpwr) and P(r|w) (avgprw), which are computed in the function logdata.m, and calculating the relevant measures. The contents of these
matrices at the end of the run are provided as output of cycle.m, and
can be accessed as follows:
[avgpwr,avgprw,avgprr] = cycle(1000);
Functions calculating
the measures are provided, and can be used as follows:
specificity(avgprw)
consistency(avgpwr)
fidelity(avgprr)
As an example, let us assume the following randomly generated
production matrix: avgpwr =
pwr =
| 0.158 | 0.203 | 0.188 | 0.451 |
| 0.016 | 0.093 | 0.779 | 0.112 |
| 0.041 | 0.013 | 0.533 | 0.413 |
| 0.128 | 0.420 | 0.360 | 0.093 |
References
De Jong, E.D. and L. Steels (2004). A Distributed Learning Algorithm
for Communication Development. Complex Systems, vol. 14, no. 4.
Contact information
For more information, see Edwin de Jong's homepage, or send email to: d e j o n g @ c s . u u . n l