Thesis (M.Sc.)--University of the Witwatersrand, Faculty of Science, School of Computer Science and Applied Mathematics, 2017. / Partially Observable Markov Decision Processes (POMDPs) provide a principled mathematical
framework allowing a robot to reason about the consequences of actions and
observations with respect to the agent's limited perception of its environment. They
allow an agent to plan and act optimally in uncertain environments. Although they
have been successfully applied to various robotic tasks, they are infamous for their high
computational cost. This thesis demonstrates the use of knowledge transfer, learned
from previous experiences, to accelerate the learning of POMDP tasks. We propose
that in order for an agent to learn to solve these tasks quicker, it must be able to generalise
from past behaviours and transfer knowledge, learned from solving multiple tasks,
between di erent circumstances. We present a method for accelerating this learning
process by learning the statistics of action choices over the lifetime of an agent, known
as action priors. Action priors specify the usefulness of actions in situations and allow
us to bias exploration, which in turn improves the performance of the learning process.
Using navigation domains, we study the degree to which transferring knowledge
between tasks in this way results in a considerable speed up in solution times.
This thesis therefore makes the following contributions. We provide an algorithm
for learning action priors from a set of approximately optimal value functions and two
approaches with which a prior knowledge over actions can be used in a POMDP context.
As such, we show that considerable gains in speed can be achieved in learning subsequent
tasks using prior knowledge rather than learning from scratch. Learning with
action priors can particularly be useful in reducing the cost of exploration in the early
stages of the learning process as the priors can act as mechanism that allows the agent
to select more useful actions given particular circumstances. Thus, we demonstrate how
the initial losses associated with unguided exploration can be alleviated through the
use of action priors which allow for safer exploration. Additionally, we illustrate that
action priors can also improve the computation speeds of learning feasible policies in a
shorter period of time. / MT2018
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:wits/oai:wiredspace.wits.ac.za:10539/24175 |
| Date | January 2017 |
| Creators | Mabena, Ntokozo |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Thesis |
| Format | Online resource (120 leaves), application/pdf |
Page generated in 0.013 seconds