General principles for paper-writing:

1) First you tell them what you are going to tell them, then you tell them, then you tell them what you just told them.

2) don't start writing text right away. Put down placeholders first, and fill placeholders with text later as results emerge

3) figures make good placeholders. arrange figures in the paper first, and with time "dress them up" with text accompanying them

4) Don't wait to write literature review until you have read all the papers - or it will never happen. Record the message of each paper inside your own, and with time you will move these records into the appropriate places to illustrate your own findings.

5) Don't be afraid to copy and paste other people's words so long as you modify them for your application.

Distributed Decision Making in a Simple Animal

alternative titles:

Robotic mollusk uses cost-benefit analysis in decision-making

A simple robot to study decision making using artificial neural networks

Liudmila S. Yafremava

University of Illinois at Urbana-Champaign

2007

Abstract

will fill-in later

Introduction

This is where I tell them what I am going to tell them.

Animal subject : Pleurobranchaea californica; where it lives, its behavior, why it is a good organism to study decision-making.

Problem : decision making. What do we know about it? How is it different in different organisms? There is a lot of info out there about this, so must narrow down on something.

The following i copied from my project plan:

I would like to investigate neuronal mechanisms of decision making in a simple animal, such as a marine slug pleurobranchaea californica. specifically, i hypothesize that simple animals, such as invertebrates, and on some level possibly higher animals as well, use distributed decision making: there is no single cell or single network, a “decision center”, responsible for negotiating the stimuli of various modalities and selecting a single motor output appropriate for the situation. i think that instead networks responsible for different behaviors all get their own appropriate stimuli from the sensors, and the behavioral decision emerges from the inhibitory and excitatory interactions between them. the question is - how?

so, in this work i would like to answer the following specific questions:

1) how do complex behavioral decisions arise from the interactions of neural networks responsible for different behaviors, such as feeding, escape, orienting and avoidance?

2) how can this complex interaction be altered by internal state, e.g. hunger, and bias the decision?

Figure

A / Bhere goes my robot pic

Proposed solutions to the problem

This I copied from my project plan:

the organism in question, the marine slug pleurobranchaea californica, is an excellent model system for studying distributed decision-making, because its neuronal pattern generators (cpgs) combine both motor neurons and command neurons. these cpgs responsible for feeding, escape swim, and turning, have been extensively studied, and are composed of a small number of identified neurons. effect of internal state on some of these networks has been researched as well. this amount of data should allow use of an integrative approach to answer the posed questions. it should be possible to use neural network simulation, electrophysiology and robotics/neuromorphic engineering as complementary tools, as they make different aspects of the investigation easier. experimentation provides data for analysis with computer and robotic simulations, while the other two suggest further experiments to improve our knowledge of the system and allow to test hypotheses otherwise unaccessible in live animal experiments, once a high-fidelity simulation has been constructed. robotic simulation is particularly useful, as it allows faster search of the parameter space, instant feedback from the model and access to realistic environmental stimuli. in contrast, computer simulation helps build the robot by working out schematics for the neural network. thankfully, much of necessary electrophysiology and computer simulation has been previously done. it is time now to use all this information to make a robotic implementation, a roboslug.

1 sentence results I don't know what they are going to be, so this is a placeholder.

methods

The following I copied from my project plan:

components of the roboslug:

the brain

consists of 3 networks: feeding (3 neurons), escape (3 neurons) and turning (4 neurons)

i will get the schematics for them later from electrophysiology papers

figure

A. i will probably modify this, because it has been previously published / B here goes picture of neuron-and-synapse circuit ... not sure if it should include the PIC

actuators

1)in the slug feeding occurs through repetitive protraction and retraction of a proboscis, a part of feeding apparatus that grabs and swallows food. it will be mimicked by a bolt that is attached to a stepper motor and is mounted on a stationary nut. as the protractor neuron fires, the motor rotates, the screw rotates with it and comes forward through the nut. i could mount a grabber on it, like those they use to get stuff out of sink drains. that will mimic biting. i don't have to mimic swallowing.

2)biomechanics of locomotion is not a concern for me in this study, only the turning: orienting or avoidance. so, i will put the robot brain on a platform mounted on 4 wheels. turning network will determine direction of motion. i don't know yet how to determine the speed of locomotion, so roboslug could slow down when it reaches food.

3)escape – is a complicated topic. i will leave it for later

internal state can be modeled with a charge on a really big capacitor: when it's full, the animal is satiated. the capacitor will normally be discharged very slowly, and as it does so, the robot will become “hungrier” and attack more dangerous food sources (need to think how to model danger). food sources will be simply modified power outlets, so that when my sewage grabber protracts, it will get in contact with the power source, and the capacitor will get charged. the robot will sense presence of food. it can be implemented in a variety of ways, either ir sensors on robot responding to the strategically placed tv remote controls, or tiny microphones detecting sound generated by computers in the lab.

I will need to implement a distinction between appetizing and noxious food somehow.

Here also goes a table with all parameters: R,C etc.

PIC?

Results

How I tested my robotic circuits

Need to compare output of my circuits with the output of real biological CPGs in the slug.

Figure - I am guessing I need to make one figure for each feeding, swimming and turning.... That's a lot of figures, and is impractical. will insert now and prune later!

A this is an old figure and will be replaced / B here goes an example of a real slug network firing pattern. Borrow from published papers with permission from the authors?

testing of swimming

testing of turning

testing of feeding

How my robotic circuits interacted:

demonstrate how a "decision" can be made without environmental stimuli by exciting one network and thereby suppressing all others.

some figures go here, don't know which.

Representation of roboslug's environment:

describe it here in some detail and explain how each network reacts to each stimulus

Decision making

present roboslug with conflicting stimuli and record from neurons, videotape behavior. Demonstrate behavioral switching depending on internal state (hunger)

Discussion

I am not ready to write this section, but will have to roughly cover the following:

1) are there any other robotic slugs?

2) are there any other decision-making robots? if yes, do they use a processor for decision-making, or a neural network?

3) usefulness of an explicit neural network approach.

4) what the roboslug allowed me to do in order to study decision-making, which was not possible in a real animal?

5) any interesting emergent properties of robotic implementation I could not foresee?

6) future directions or application of the robot?

References:

will start inserting next week.

conclusions ???

i

------

once all networks are implemented and the internal state capacitor is connected to them as described in electrophysiology papers, i will allow the robot to explore environment, the lab, which will contain tasty and noxious food, as well as some predators. the neural networks will respond to stimuli and make “decisions” whether to approach a particular stimulus or avoid it, and when to escape. i will monitor the patterns of excitation within the neural networks, perform lesion experiments to knock out neurons and determine which are important for decisions. i will aim at determining constraints, the ranges of parameters in neural networks, that result in optimum decisions.

------

webb 2000 lsagsvbsjty