Supplementary Materials for "Generalized Bargaining Mechanisms" (Submitted to AAAI 2023 Senior Member Presentations Track)
This folder contains supplementary materials as well as all the code implementing the algorithms described in the paper and all datasets used.
- This guides shows how to reproduce every result in the paper and more.
- The data is already publicly available in the GENIUS platform.
- All the content of our supplementary materials will be made publicly available on Github in case of acceptance. Moreover, we will contribute our algorithms back the negotiation platform we use (NegMAS) as a pull request and if accepted will be available directly from the platform (which is community driven) in the future.
supplementary.pdf
Contains proofs and further experimental results and discussions.code
Implementation of algorithms, negative cases, and experiments.negmas
A fork of the NegMAS platform including an implementation of a simplified version of GBMs GBM(1,1,other,.,.,.) undernegmas/negmas/gb
using which AOP, SAOP and TAU are implemented.
data
All the negotiation scenarios used in the experiments.
-
algorithms
Contains implementations of the TAU protocol and SCS protocol. Implementations are modular and use a generalized version of the protocol description method introduced in the paper (Subsections Negotiation Protocol and Negotiation Strategy). -
experiments
All code needed to run the experiments reported in the paper and generate the figures and tables in it.-
experiment1.sh
Runs experiment1
and generatese1.csv
with all raw results of the experiment. -
experiment2.sh
Runs experiment2
and generatese2.csv
with all raw results of the experiment. -
analyze1.sh
Runs analysis of the first experiment and generates Table 1 of the paper in latex format. It usese1.csv
generated byexperiment1.sh
. To just run the analysis on the data already in the data folder, changee1.csv
in this script toe1.csv
. -
analyze2.sh
Runs analysis of the second experiment and generates Fig. 1 of the paper in latex format. It usese2.csv
generated byexperiment2.sh
. To just run the analysis on the data already in the data folder, changee2.csv
in this script toe2.csv
. -
experiment.py
The actual code used to run all experiments (Called by the experiment shell scripts above). It is a command line utility that can be called with different arguments to run different experiments. -
eval1.py
Used to analyze experiment1
(flexible and can be adjusted to do other evaluations as well not reported in the paper). -
eval2.py
Used to analyze experiment2
(flexible and can be adjusted to do other evaluations as well not reported in the paper). -
utils.py
Utility methods used by other python scripts -
helpers
Helper scripts containingclean.py
which is used in the second experiment to clean out extra runs if any and runs without controlled reserved value. In principle, it should not be needed. -
tests
Contains scripts for testing AOP and TAU strategies on specific conditions.-
testaop.py
Tests an AOP strategy (default isMiCRO
). You can run three simple experiments with it:-
e1
(incomplete
) shows thatMiCRO
andAtlas3
are not exactly complete using the tinyNiceOrDie
scenario (other strategies may vary). -
e2
(all-rational
) shows success on the same domain when all outcomes are rational. -
e3
(none-rational
) shows expected disagreement on the same domain when no outcomes are rational. -
e4
(small
) runs the negotiation a relatively small domain (Outfit
) -
e5
(large
) runs the negotiation on a relatively large domain (Kitchen
) -
e6
(difficult
) runs the negotiation on a relatively large domain (Outfit
) with larger reserved values (more difficult agreement) -
free
can be used to run on any domain.
-
-
testtau.py
Tests the proposed SCS strategy on few example cases.-
e1
(complete
) shows that TAU($\infty, \infty$ ) is not exactly complete using the tinyNiceOrDie
scenario (TAU($\infty, 0$ ) does get an agreement). -
e2
(all-rational
) shows success on the same domain when all outcomes are rational. -
e3
(none-rational
) shows expected disagreement on the same domain when no outcomes are rational. -
e4
(small
) runs the negotiation a relatively small domain (Outfit
) -
e5
(large
) runs the negotiation on a relatively large domain (Kitchen
) -
e6
(difficult
) runs the negotiation on a relatively large domain (Outfit
) with larger reserved values (more difficult agreement) -
free
can be used to run on any domain.
-
-
-
-
domains
All the negotiation domains used for the experiments reported in the papers. These are all the domains of the ANAC 2013 competition and they can also be found on the GENIUS platform. Each domain folder consists of the following files:{domain-name}.xml
describes the outcome space in GENIUS XML format.{domain-name}1.xml
describes the utility function of first negotiator (GENIUS XML format).{domain-name}2.xml
describes the utility function of second negotiator (GENIUS XML format).stats.json
Statistics of the negotiation scenario like opposition, welfare of the Nash Bargaining Solution, etcinfo.json
More information about the negotiation scenario for different choices of reserved values including the fraction of rational outcomesf_rational
.
Note that folder names prefix each domain with the size of its outcome space.
-
results
Raw results of all negotiations in experiments (e1
for experiment 1 ande2
for experiment 2) as reported in the paper.stats
Contains the results of all statistical tests reported in the paper (as well as others not reported due to lack of space).
- Python
3.9
or later (tested on MacOS with Python 3.10.5). - Java
18
(tested on MacOS with OpenJDK 18 2022-03-22 build 18+36-2087).- only needed to run state-of-the-art negotiators (and Nice Tit for Tat)
- negmas
0.9.2
(our forked version included in this code repository and installed when following the instructions in the following section). - genius-bridge
v0.2.4
(installed when following the instructions in the following section).- only needed to run state-of-the-art negotiators (and Nice Tit for Tat)
All experiments were done on MacOS 12.4 on a MacBook Pro with Apple M1 Pro processor and 16GB of RAM but should be reproducible on other operating systems as we only use OS-independent libraries in this work.
Please use your platform's preferred method to install Python 3.9+ and Java 18+.
To install other requirements run the following command within the code
folder:
pip install -r requirement.txt
negmas genius-setup
Note that the later command will download the negmas-genius
bridge and install in as:
$HOME/negmas/files/geniusbridge.jar
We assume that this is run within a virtual environment (as always recommended).
Assumes that you installed requirements
To run the first experiment and generate Table 1
, run the following within code/experiments
:
> java -jar $HOME/negmas/files/geniusbridge.jar --silent --no-logs&
> experiment1.sh
> eval1.sh
To run the second experiment and generate Fig. 1
, run the following within code/experiments
:
> java -jar $HOME/negmas/files/geniusbridge.jar --silent --no-logs&
> experiment2.sh
> eval2.sh
-
These experiments may take a LONG time mostly because of the relatively slow speed of some state-of-the-art algorithms.
-
In case, the previous method did not work, it may be because of an issue with the NegMAS-GENIUS bridge. To investigate the issue you will need two terminals. In the first terminal run:
> java -jar $HOME/negmas/files/geniusbridge.jar --verbose --debug --no-logs
Now, In the second terminal run the remaining two lines. For example for the second experiment:
> experiment2.sh > eval2.sh
You can use code/experiments/tests/test*.py
to run single negotiations and test properties of TAU and AOP on different situations.
To run one of the five predefined examples given in the description above just pass its name. For example, to run the negotiation that shows that MiCRO is not complete, run:
> python testaop.py e1
or
> python testaop.py incomplete
To change the strategy pass it using --strategy
, for example:
> python testaop.py e1 --strategy=Atlas3
To change the domain and run a free negotiation use the free option
> python testaop.py free --strategy=micro --domain=<path to the domain folder>
You can run similar negotiations using the TAU protocol
> python testtau.py e1
By default TAU will be used, but you can change --beta
. For example to show that TAUNR is not complete you can run:
> python testtau.py e1 --beta=-1
Passing --beta
less than zero will be interpreted as