Loading…
CAA2016 has ended
Program

Friday, April 1 • 09:45 - 10:10
S14-04 Reinforcement learning for decision making in agent-based models

Sign up or log in to save this to your schedule and see who's attending!

Reinforcement learning for decision making in agent-based models

Jean-Marc Montanier, Xavier Rubio-Campillo

Abstract
Understanding the decision-making processes within past societies is a
challenging aspect for archaeological research. In order to validate our
understanding of these behaviours, an ideal workflow would be to simulate the
behaviours that we imagine are correct. We can then observe, if the imagined
behaviours reproduce the evidence collected from the terrain.

ABM has proved to be an efficient approach toward the realisation of this
vision. However, many of the phenomenons we wish to study, require the
adaptation of the agents to the context they live. For example, it would be
interesting to study the behaviors of gatherer agents in front of changing
environments. To face this type of situations, there is a clear need for
adaptation abilities. Unfortunately, most of the current behavioural
architectures used in ABMs do not let the agent adapt continuously its
behaviour so as to fit the environment, thus restricting the modeling
possibilities.

Reinforcement learning algorithms have been proposed to address learning
problems within multi-agent settings. Once applied to past-societies models,
this learning method faces two main challenges: each agent observes only part
of the world and the number of states and actions an agent can face is
potentially extremely large. Similar challenges have been previously been
encountered by researcher applying reinforcement learning methods to
multi-robots scenarios. Within this article we aim to present which of the
solutions previously developed can be applied to create models of past
societies.

Moreover, the use of a UCT architecture has been previously proposed to address
the challenge of learning in past-societies models. We will draw a comparison
between reinforcement learning and UCT approaches. This comparisons will
highlight the difficulties each approach is facing, specifically for an
application to past-societies models


Friday April 1, 2016 09:45 - 10:10
Domus Bibliotheca

Attendees (1)