Adaptive Representations for Parameter Optimization
- Daniel Ashlock
Selecting parameters that permit a model to best fit a set of data is a
very old problem. For many models there are rapid, optimal regression techniques.
Consider a man searching under the streetlight for a key he lost in the
park because the light is better - even though the key was not lost near
the streetlight. Our choice of models is influenced by the availability
of good regression algorithms, even when they lead to less-than-appropriate
models. Evolutionary computation is an example of a relatively low-speed
but extremely general model fitting technique. In this talk non-adaptive
and adaptive techniques for fitting model parameters with evolutionary computation
are defined, explained, and contrasted. This talk touches on the issue of
representation in evolutionary computation; that is, choosing good ways
to represent problems for computer solution.
Evaluating Epoetin dosing strategies using observational
longitudinal data
- Cecilia Cotton
Epoetin is commonly used to treat anemia in Chronic Kidney Disease and
End Stage Renal Disease subjects undergoing dialysis, however, there is
considerable uncertainty as what level of hemoglobin or hematocrit should
be targeted in these subjects. In order to address this problem we treat
epoetin dosing strategies as a type of dynamic treatment regimen. We present
a methodology for comparing the causal effects of multiple treatment regimens
on survival based on observational data. This problem is complicated by
the fact that depending on the regimen definitions, subjects may have been
adherent to multiple regimens at the same time. We present a methodology
in which each subject is cloned (or replicated) and contributes follow-
up data to each regimen to which they were continuously adherent before
being artificially censored. We provide an inverse probability weighted
log-rank test with variance estimate that can be used to compare survival
under two regimens. For comparing multiple regimens we propose several marginal
structural Cox proportional hazards models with robust variance estimation.
The methods are illustrated through simulations and applied in an analysis
comparing epoetin dosing regimens in a cohort of 33,873 adult hemodialysis
patients from the United States Renal Data System.
Swarm Intelligence for Unmanned Aerial Vehicles
- David Howden
Wildfires are destructive conflagrations that occur in areas of wilderness
and their remote location serves as a barrier to rapid detection or response.
Due to their inaccessibility, these fires can grow to insuppressible proportions
and not only cause significant economic damage to an area, but also endanger
the lives of communities and their fire fighters. Fast and effective detection
is a key factor in bushfire fighting. Accurate knowledge of the status of
a bushfire is indispensable for enabling accurate fire prediction modelling,
maintaining the safety of fire fighting crews, and allowing efforts to be
focused on areas of the highest risk such as urban areas with strong human
presence. This problem is ideal for the application of multiple unmanned
aerial vehicles. In this presentation, a swarm intelligence approach to
exhaustive and continuous surveillance of large areas is introduced. Using
a pheromone inspired technique, landmarks and features can be assigned priorities
relative to their value or risk, either on deployment or dynamically in
reaction to observations. The presented approach is fully distributed, resilient
against loss, and not reliant on cooperative decision making or long range
communication.
Bio-inspired complex systems engineering
- Taras Kowaliw
The use of complex systems in engineering design promises to open new areas
of productivity. We start with a brief description of some of the strengths
and weaknesses of using evolutionary computation for design. We next discuss
artificial development, the use of a growth stage in the optimization process.
Rather than directly specify a design, we aim to grow it. Artificial development
is a potential means of incorporating several biologically-motivated design
metaphors, ones that might generate designs of greater complexity, robustness,
and resilience. Two concrete examples are presented: firstly, the design
of forms in the domain of structural engineering, where we recover a form
of artificial polymorphism; secondly, the generation of soft-bodied virtual
robots. Finally, we discuss potential applications to the automation of
synthetic biology.
A Clinical Trial Design for Constructing and Evaluating
Individualized Real-Time Treatment Policies
- Susan Murphy
Mobile devices, including mobile phones, are increasingly used to both
passively and actively collect patient symptoms, where the patient is, who
the patient is with, level of social activity. At the same time, mobile
devices are beginning to be used to deliver a variety of real-time behavioral
interventions (motivational assistance, cognitive assistance, suggestions
concerning social interactions). However only a few researchers have begun
to use the real-time patient information to adapt and re-adapt the behavioral
interventions to the patient. And for the most part, this adaptation is
primarily based on behavioral theory, clinical experience and expert opinion.
Data-based evidence is, at best, indirectly used in this process. In this
talk we sketch out the outline for, and solicit feedback on, a new clinical
trial design for the purpose of providing/using patient data to inform the
development of Individualized Real-Time Treatment Policies.
Generating candidate adaptive dosing
strategies through simulation and g-estimation
- Ben Rich
Warfarin, a commonly prescribed oral anticoagulant, is a drug which requires
adaptive individualized dosing. While much is known about the properties
of the drug, an optimal dynamic dosing strategy has remained elusive. We
propose the use of a realistic pharmacokinetic-pharmacodynamic model to
generate simulated patient data, to which g-estimation is applied to generate
a candidate dynamic strategy. Using the same data-generating model, different
strategies can be compared on future subjects drawn from the same population
or a different one. Our findings suggest that despite partial model misspecification
this methodology can lead to a dosing strategy which performs well both
within and across populations with varying pharmacokinetic and pharmacodynamic
characteristics.
Statistical methods for comparing
adaptive treatment strategies in SMART designs with time-to-event endpoints
- Abdus S. Wahed
Adaptive treatment strategies (ATSs; aka dynamic treatment regimes) are
individually tailored treatments, with treatment type and/or dosage changing
according to patient’s intermediate response at different stages of
therapy. Availability of multiple treatment options at each stage and possibility
of variable intermediate responses to these treatments may result in many
different adaptive treatment strategies. Sequential multiple assignment
randomization trials (SMARTs) are often used to study ATSs in the treatment
of cancer, leukemia, depression, and AIDS. Frequently the goal is to compare
multiple strategies based on time-to-event outcome such as overall survival.
In this talk we will discuss some semi-parametric approaches to compare
two-stage ATSs based on time-to-event data collected from SMART designs.
Specifically, traditional methods of estimation such as Empirical CDF, Kaplan-Meier,
and Nelso-Aalen, and test of hypothesis such as Log-Rank test will be adapted
using inverse-probability-weighted methods to be applicable to this specific
trial design. Application of these methods will be demonstrated by applying
to a cancer clinical trial dataset.
Back to top
