![](bg2.png)
comes at a cost of some inaccessability for interested
newcomers. The aim of these frameworks is to fa-
cilitate multi-agent applications such as logistics and
online trading (Rogers et al., 2007), rather than ex-
plicitly to explore large-scale distributed cognition in-
volving multiple authors. These frameworks provide
architectures that make assumptions about how in-
formation should be represented and communicated,
how systems should be organised and what their re-
sponsibilities are. It is our belief that fewer require-
ments and responsibilities placed upon mind authors
may facilitate more participation.
A major goal of the W2M project is to minimise
the barrier to entry for researchers, students and ca-
sual programmers to create and test their minds and
worlds freely (Walshe et al., 2004). Authors are free
to define the world state and action representations, as
well as the interactions between sub-minds as they see
fit. In our implementation we have selected a simple
scheme defined by an XML-based markup language
where minds return a suggested action in response to
a getaction message containing the current state ob-
servable by the agent in the world, and worlds respond
to getstate and takeaction methods.
It is noted in (Bryson, 2007) that none of the lead-
ing architectures documented in an earlier review of
action selection work were still actively used seven
years later. Perhaps this suggests that, rather than
re-inventing architectures in an attempt to create in-
telligence, what might really be needed is simply a
large set of agents implementing various behaviours
and problem-solving strategies so that, while no one
person need understand every element of the whole
mind, it is collectively created by a large community.
Collaboration and sharing exists within artificial
intelligence research; for example, there are websites
which serve as repositories for machine learning code
(Kantrowitz, 2009) and training datasets (Asuncion
and Newman, 2009). These repositories are useful,
but the steps required to install or adapt an existing
solution differ in each instance and there is little con-
sistency in the types of programs and interfaces pro-
vided. Users must download the code and often patch
it to compile on their own machine, and must adapt
the program to suit the interface and problem structure
they wish to solve, if indeed the program addresses
the chosen problem.
RoboCup (Visser and Burkhard, 2007) and the
DARPA Challenge (Rouff, 2007) are closer to what
we wish to achieve, but the problem domains are spe-
cific and there is no clear way to build and share hy-
brid minds. These projects focus on competition but
not explicitly on re-use. Nevertheless the significant
interest in the development of agents indicates that
collaborative approaches to A.I. research can be suc-
cessful.
There are limits on what one researcher or even
one lab can do. If we succeed in building a fast, us-
able platform which can be used to build diverseprob-
lem worlds and create minds composed of a multitude
of subminds - if multi-author collaboration could be
made easier - then the interest generated and subse-
quent studies could have a positive impact on A.I. re-
search.
3 WORLD-WIDE MIND “V2.0”
We outline here some of the recent changes that may
realise the vision of the World-Wide Mind project and
make it usable across the world.
3.1 Speed
The initial design of the platform (Humphrys, 2001)
embodied minds and worlds as web services, operat-
ing over the internet by sending messages over the
HTTP protocol. This enabled hybrid minds to be
built and evaluated (Walshe et al., 2004), but was im-
practically slow for large hybrids. When the system
was used as the basis for undergraduate assignments
in third year A.I. courses, the majority of submitted
minds were monolithic programs which did not seek
the advice of other minds to select actions.
As the primary objective of the project is for the
platform to be used widely by a large number of au-
thors, improving the latency and throughput of mes-
sages was a major concern.
3.1.1 Fast Communication Protocol
The implementation of minds and worlds as web
services afforded simplicity and transparency in dis-
tributing minds and worlds online. However, there is
a time penalty to be paid for the use of a web applica-
tion server (Apache Tomcat) and the HTTP protocol
to wrap messages.
To avoid these bottlenecks, the web application
server was replaced by a custom TCP-based server
which reads and writes XML messages across the net-
work as length-prefixed strings. Similarly, while one
goal of the platform is to be tolerant of errors in the
generated XML, we have optimised the case where
the XML is well-formed (since it will normally be
machine-generated).
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
660