SPRING 07 ARCHIVE Homework

Homework T2
tournament controller

Homework #9
ant trails

Possible solution:
hw09_soln.m

Some take-home messages:

• Trial marking/following can improve foraging efficiency, but the implementation is non-trivial.
• Time-varying trail deposition (exponential decay) can be useful in this context (e.g., motor.trail = 0.9 * motor.trail).
• Some sort of 'timeout' mechanism is useful to escape from lingering pheromone trails after food has been depleted.
• The ant's motor strategy in the absence of a signal (no food, no trail) is important in optimizing its ability to find remote food patches.

Homework #8
active sensing

Possible solution:
hw08_soln.m

Some take-home messages:

• Active sensing organisms incur a metabolic cost associated with generating the probe energy.
• The cost scales steeply with sensing range; it's expensive to 'light up' the entire world.
• Active sensing organisms optimize the probe energy by controlling its intensity, timing, and directionality.

Homework #7
honeybee foraging

Possible solution:
hw07_soln.m

Some take-home messages:

• "Learning from experience" is an important component of adaptive behavior in animals and animats
• Single-trial learning of color-reward associations is possible when reward distributions for the two colors are well separated.
• Deterministic control of turn angle based on local color signals can give rise to global foraging patterns that suggest the bot might have 'preferred' routes or retain memory of previously visited locations. But, having programmed the bots ourselves, we know this is not the case. We must be careful when interpretting animal behavior!

Homework T1
tournament controller

Homework #5
restricted search

Possible solution:
hw05a_soln.m (edge following)

Some take-home messages:

• organisms can improve resource acquisition performance by biasing their search to regions that are more likely to be profitable (e.g. edges in hw05a, regions where food has been found recently in hw05b)
• most neurobiological research involves analysis and reverse engineering of biological systems
• the animat approach focuses on synthesis and forward engineering of biologically inspired systems
• Braitenberg's principle of uphill analysis and downhill synthesis states that designing a complex system from scratch is generally less difficult than taking one apart to figure out how it works

Homework #4
competition, light, whoppers

Possible solutions:
hw04b_soln.m (uses light cues)
hw04c_avoid.m (avoid whoppers by going around them)
hw04c_escape.m (avoid whoppers using 'escape' behavior)
hw04c_escape2.m (escape based on time since last meal)

Some take-home messages:

• competition changes the fitness landscape, speed becomes important, 'the early bird catches the worm'
• predictive informational cues (like light in hw04b) can help an organism get to resources more quickly
• nonlinear sensor processing: more is not always better

Homework #3
Braitenberg vehicles

Possible solutions:
hw03_hard.m (hard transition between approach/avoid)
hw03_soft.m (soft transition between approach/avoid)

Some take-home messages:

• bilaterally paired sensors and actuators enable a broad range of positive and negative taxis behaviors
• a rich repertoire of behavioral responses can be achieved using CROSSED/UNCROSSED and EXCITATORY/INHIBITORY connections between sensors and actuators
• taxis (directed locomotor responses) can outperform kinesis (non-directed responses) in resource acquistion tasks

Homework #2
single-sensor strategies

Possible solution: hw02_soln.m

Some take-home messages:

• simple, non-oriented (kinesis) strategies can result in spatial clustering of animats in their environment
• the efficacy of a control strategy depends on the structure of the environment; a good strategy in one environment may not work well in another
• understanding the 'brain' requires understanding the environment in which it operates

Homework #1
pacman foraging

Possible solutions: hw01_soln1.m, hw01_soln2.m
Finite-state machine: hw01_fsm.m, hw01_fsm_evol.m
Neural net (5 binary neurons): hw01_nn.m,

Some take-home messages:

• reactive control (direct sensor-motor mapping) is insufficent to solve even relatively simple tasks
• in general, animals and animats select actions based on a combination of sensory input and internal state
• the same behavior can result from different internal control structures (if..then...statements, finite-state machine tables, neural networks, ....); observing the agent's behavior is not sufficent to predict what's 'inside the box'
• simulated evolution was able to find a controller that was more efficient (required fewer moves) than any of the controllers we designed by hand (see hw01_fsm_evol.m)