• Classified by Topic • Classified by Publication Type • Sorted by Date • Sorted by First Author Last Name • Classified by Funding Source •
Program Embeddings for Rapid Mechanism Evaluation.
Sai Kiran Narayanaswami, David Fridovich-Keil, Swarat Chaudhuri,
and Peter Stone.
In ICRA Workshop on Multi-Robot Learning, May 2023.
[PDF]1.3MB [poster.pdf]916.2kB
Mechanisms such as auctions, voting and traffic control systems incentivizeagents in multiagent systems in order to achieve or optimize certain globaloutcomes such as safety, productivity, and welfare. Designing mechanisms in theform of programs would bring several benefits such as interpretability,transparency and verifiability. However, program synthesis for finding suitablemechanisms from a search space represented by programs leads to computationallyexpensive two-level optimization problems, where mechanisms need to be evaluatedin an inner loop, which entails learning agent responses for each mechanism.Multi-Task Learning (MTL) approaches efficiently learn parameterized agentstrategies that can be used to act in various tasks, here mechanisms defined byprograms. Such multi-task agent policies allow for rapid evaluation of mechanismperformance for a given program by bypassing the need to learn agent responsesfrom scratch, thus allowing for efficient evaluation of mechanisms. In this work,we use learned program representations embeddings to provide suitable taskcontexts that make MTL feasible with combinatorially large programmatic searchspaces of mechanisms. We demonstrate experimentally that program embeddingsgenerated by an off-the-shelf Code2Vec model are sufficient for reconstructingmatrix games based on only a programmatic description. The embeddings are alsoable to serve as task context to guide a set of agent strategies to actnear-optimally in mechanisms from the search space.
@InProceedings{nskiran_ICRA2023WMRL,
author = {Sai Kiran Narayanaswami and David Fridovich-Keil and Swarat Chaudhuri and Peter Stone},
title = {Program Embeddings for Rapid Mechanism Evaluation},
booktitle = {ICRA Workshop on Multi-Robot Learning},
year = {2023},
month = {May},
location = {London, UK},
abstract = {Mechanisms such as auctions, voting and traffic control systems incentivize
agents in multiagent systems in order to achieve or optimize certain global
outcomes such as safety, productivity, and welfare. Designing mechanisms in the
form of programs would bring several benefits such as interpretability,
transparency and verifiability. However, program synthesis for finding suitable
mechanisms from a search space represented by programs leads to computationally
expensive two-level optimization problems, where mechanisms need to be evaluated
in an inner loop, which entails learning agent responses for each mechanism.
Multi-Task Learning (MTL) approaches efficiently learn parameterized agent
strategies that can be used to act in various tasks, here mechanisms defined by
programs. Such multi-task agent policies allow for rapid evaluation of mechanism
performance for a given program by bypassing the need to learn agent responses
from scratch, thus allowing for efficient evaluation of mechanisms. In this work,
we use learned program representations embeddings to provide suitable task
contexts that make MTL feasible with combinatorially large programmatic search
spaces of mechanisms. We demonstrate experimentally that program embeddings
generated by an off-the-shelf Code2Vec model are sufficient for reconstructing
matrix games based on only a programmatic description. The embeddings are also
able to serve as task context to guide a set of agent strategies to act
near-optimally in mechanisms from the search space.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Tue Nov 19, 2024 10:24:46