Skip to content

Latest commit

 

History

History
59 lines (36 loc) · 3.67 KB

wait_im_confused.md

File metadata and controls

59 lines (36 loc) · 3.67 KB
Raw

Wait! I'm confused by the several BT packages here!

First it is recommended that you read two papers:

  1. Simulation Results on Selector Adaptation in Behavior Trees
  2. "Hidden Markov Models derived from Behavior Trees".

for context, terminology, notation, and examples.

OK, let's start with a little chronology:

  1. In 2013, Marzinotto et al., 2014, a great group out of KTH in Sweden, got to work on BTs.

  2. They released a javascript package and a python version shortly after. See https://github.com/behavior3.

  3. Blake Hannaford modified the python version (behavior3py) so that BT selector nodes could learn from experience: Adaptive BTs, as documented in the report: Simulation Results on Selector Adaptation in Behavior Trees. These mods are in the branch abt_dev

  4. BH worked on a new class of BTs, Augmented BTs, in which each node emits observations drawn from a probability distribution, and has a defined probability of success on each tick. These were analytically linked to Hidden Markov Models in the report "Hidden Markov Models derived from Behavior Trees".

  5. In the future, someone could combine the classes so that a single derived set of classes supports Adaptive BTs and Augmented BTs.

Classes:

Adaptive BTs (abts)

The purpose of Adaptive BTs is to learn from experience. For example, if the third child of a Selector node with 5 children has the highest probability of success (as learned over multiple tics) then it should be tried first by the selector node instead of after the first two nodes fail.

Adaptive BTs are built on the behavior3py module with minor modifications[1]. The key functionality of adaptive BTs is the selector (Priority) node which chooses the most successful child first (as opposed to a fixed order). This is the SmrtSel00(), SmrtSel01(), etc. classes (featuring varieties of the smart selection algorithm). SmrtSelxx() are located in the adaptive-b3 project main branch.

Augmented BTs (ABTs)

The purpose of Augmented BTs is to

  1. Study the relationship between Behavior Trees and HMMs
  2. Track evolution of a BT in spite of noisy measurements.
  3. Model behavioral systems with stochastic behavior and noisy measurements.

Augmented BTs refer to BTs in which each leaf (class aug_leaf()) inherits the behavior3py.Action() class and adds

self.Ps   # the probability the node will return `b3.SUCCESS' (a fixed param)
self.Pf = 1.0-self.Ps
self.Obs = np.zeros(NSYMBOLS)   # probability to emit each observation symbol in that state.

Sadly, instead of a proper import structure, b3 is currently symlinked or hard-copied into the ABT repository. If the abt_dev branch of behavior3py [1] is updated, recopy the entire b3 directory from it into ABT/ .

how the modules include each other

Augmented BTs are analytically equivalent to HMMs (e.g. there is exactly one HMM for each ABT). HMM algorithms are applied to outputs from executed ABTs in various "top level" test software.

Notes

[1] b3 library (a fork of behavior3py, as modified in the abt_dev branch