treez
A collection of useful data structures and algorithms
current implementations: segment tree, rb tree, reverse auto gradient differentiator, indexed tree, sarsa Q-Learning
work in progress: variants of mcts-related search, autograd vectorization
note: this is work in progress and interfaces may change from version to version
segment tree
implementation: array based
todo: generic type
notes: for static use after initialization
let mut segments = vec![];
for i in 0..10 {
let n = (i*5, 5*i+5, i); segments.push( n );
}
let t : treez::seg::TreeSeg< i32, i32 > = treez::seg::TreeSeg::init( segments.as_slice() );
let query_segs: HashSet<_> = t.get_segs_from_bound( (15,20) ).iter().cloned().collect();
let check: HashSet<_> = [ 2, 3, 4 ].iter().cloned().collect();
println!( "query segs: {:?}", query_segs );
assert!( check.intersection(&query_segs).count() == check.len() );
red black tree
implementation: array based, threshold compaction, minimal heap allocation
todo: optimize internal representation and operations
notes: comparable performance to BTreeMap
let mut t : treez::rb::TreeRb< isize, isize > = treez::rb::TreeRb::new();
for i in 0..nums.len() {
let r = nums[i];
t.insert( r, i as isize );
}
for i in 0..nums.len() {
let r = nums[i];
let v = t.remove( &r ).expect( "remove unsuccessful" );
}
reverse automatic gradient differentiation
implementation: computes the gradient dy/dxi from reverse computation graph and caches results for computed dy/dxi where xi preceeds y in the forward computation graph
todo: add more test coverage, tweek to more ergonomic interface, optimization and parallelism
let mut c: autograd::Context = Default::default();
let buf = {
let mut x = c.init_var( &[6f64, 5f64] );
let mut y = c.init_var( &[7f64, 3f64] );
let mut z = c.init_op( autograd::OpType::Mul, & mut [ & mut x, & mut y ] );
let mut a = c.init_var( &[3f64, 8f64] );
let mut b = c.init_op( autograd::OpType::Add, & mut [ & mut z, & mut a ] );
vec![ x, y, z, a, b ]
};
let var_ids = c.fwd_pass( buf ).unwrap();
let mut var_map = HashMap::new();
for i in [ "x", "y", "z", "a", "b" ].iter().zip( var_ids ) {
var_map.insert( i.0, i.1 );
}
{
let mut var_grad = HashMap::new();
let b_id = *var_map.get(&"b").unwrap();
for i in var_map.iter() {
let grad = c.compute_grad( b_id, *i.1 ).unwrap();
var_grad.insert( *i.0, grad );
}
assert_eq!( c.get_var(*var_map.get(&"z").unwrap()).unwrap()._val, &[ 42f64, 15f64 ] );
assert_eq!( c.get_var(*var_map.get(&"x").unwrap()).unwrap()._val, &[ 6f64, 5f64 ] );
assert_eq!( c.get_var(*var_map.get(&"y").unwrap()).unwrap()._val, &[ 7f64, 3f64 ] );
assert_eq!( c.get_var(*var_map.get(&"b").unwrap()).unwrap()._val, &[ 45f64, 23f64 ] );
assert_eq!( c.get_var(*var_map.get(&"a").unwrap()).unwrap()._val, &[ 3f64, 8f64 ] );
assert_eq!( var_grad.get(&"z").unwrap(), &[ 1f64, 1f64 ] );
assert_eq!( var_grad.get(&"x").unwrap(), &[ 7f64, 3f64 ] );
assert_eq!( var_grad.get(&"y").unwrap(), &[ 6f64, 5f64 ] );
assert_eq!( var_grad.get(&"b").unwrap(), &[ 1f64, 1f64 ] );
assert_eq!( var_grad.get(&"a").unwrap(), &[ 1f64, 1f64 ] );
}
{
let z_id = *var_map.get(&"z").unwrap();
let a_id = *var_map.get(&"a").unwrap();
let grad = c.compute_grad( z_id, a_id ).unwrap();
assert_eq!( &grad[..], &[ 0f64, 0f64 ] );
}
prefix sum tree
implementation: array based
todo: support generic commutative operation
let mut t = treez::prefix::TreePrefix< isize >::init(16);
t.set(0, 5);
t.set(1, 7);
t.set(10, 4);
assert_eq!( t.get_interval(0, 16), 16isize );
assert_eq!( t.get_interval(10, 11), 4isize );
assert_eq!( t.get_interval(1, 11), 11isize );
t.set(1, 9);
assert_eq!( t.get_interval(1, 2), 9isize );
assert_eq!( t.get_interval(1, 11), 13isize );
assert_eq!( t.get_interval_start( 2 ), 14isize );
assert_eq!( t.get_interval_start( 11 ), 18isize );
t.add( 0, 1);
assert_eq!( t.get_interval_start( 2 ), 15isize );
assert_eq!( t.get_interval_start( 11 ), 19isize );
sarsa policy search
implementation: using eligibility trace, configurable reward decay and rollout factors, Q-learning based, basic thread parallel implementation
notes: This is an implementation attempt based on readings from various sources such as Reinforcement Learning by Sutton et al.
todo: switch to fine grained parallelism
use self::treez::sarsa;
use std::f64;
#[derive(Clone)]
pub struct GameGridWorld {
_dim: ( usize, usize ),
_start: ( usize, usize ),
_end: ( usize, usize ),
_obstacles: Vec< (usize,usize) >,
_play_path: Vec< ( State, Action ) >,
}
#[derive(Eq, PartialEq, Clone, Hash, Debug)]
pub enum Action {
UP,
DOWN,
LEFT,
RIGHT,
NONE,
}
#[derive(Eq, PartialEq, Clone, Hash, Debug)]
pub struct State( usize, usize );
impl sarsa::Game< State, Action > for GameGridWorld {
fn gen_initial_state( & mut self ) -> State {
self._play_path.clear();
let s = State( self._start.0, self._start.1 );
s
}
fn gen_possible_actions( & mut self, s: & State ) -> Vec< Action > {
let mut actions = [ true; 4 ];
for x in self._obstacles.iter() {
if ( s.0 >= self._dim.0 - 1 ) || ( s.0 + 1 == x.0 ) && ( s.1 == x.1 ) {
actions[0] = false;
}
if ( s.0 == 0 ) || ( ( s.0 - 1 == x.0 ) && ( s.1 == x.1 ) ) {
actions[1] = false;
}
if ( s.1 >= self._dim.1 - 1 ) || ( ( s.0 == x.0 ) && ( s.1 + 1 == x.1 ) ) {
actions[2] = false;
}
if ( s.1 == 0 ) || ( ( s.0 == x.0 ) && ( s.1 - 1 == x.1 ) ) {
actions[3] = false;
}
}
if s.0 >= self._dim.0 - 1 {
actions[0] = false;
}
if s.0 == 0 {
actions[1] = false;
}
if s.1 >= self._dim.1 - 1 {
actions[2] = false;
}
if s.1 == 0 {
actions[3] = false;
}
let v : Vec< Action > = actions.iter().zip( &[ Action::RIGHT, Action::LEFT, Action::UP, Action::DOWN ] ).filter_map( |x| if *x.0 == true { Some( x.1.clone() ) } else { None } ).collect();
v
}
fn do_action( & mut self, s: & State, a: & Action ) -> ( sarsa::Reward, State ) {
let s_next = match *a {
Action::LEFT => {
(s.0 - 1, s.1)
},
Action::RIGHT => {
(s.0 + 1, s.1)
},
Action::UP => {
(s.0, s.1 + 1)
},
Action::DOWN => {
(s.0, s.1 - 1)
},
_ => { panic!{}; },
};
self._play_path.push( ( s.clone(), a.clone() ) );
let r = if s_next == self._end {
sarsa::Reward(1.)
} else {
sarsa::Reward(0.)
};
( r, State( s_next.0, s_next.1 ) )
}
fn is_state_terminal( & mut self, s: & State ) -> bool {
self._end.0 == s.0 && self._end.1 == s.1
}
fn get_state_history( & self ) -> Vec< ( State, Action ) > {
self._play_path.clone()
}
fn set_state_history( & mut self, h: & [ (State, Action) ] ) {
self._play_path = h.to_vec();
}
}
fn main() {
let mut game = GameGridWorld {
_dim: (5,5),
_start: (4, 0),
_end: (0, 0),
_obstacles: vec![ (1, 0), (1, 1), (1, 2) ],
_play_path: vec![],
};
let sc = sarsa::SearchCriteria {
_lambda: 0.99,
_gamma: 0.9,
_alpha: 0.03,
_stop_limit: sarsa::StopCondition::TimeMicro( 10_000_000.0 ), _policy_select_method: sarsa::PolicySelectMethod::Softmax,
};
let ( _policy_map, policy_normalized, expectation, iter ) = sarsa::search( & sc, & mut game ).unwrap();
for i in (0..game._dim.1).rev() {
let mut v = vec![];
for j in 0..game._dim.0 {
let actions_percentage = match policy_normalized.get( &State( j, i ) ) {
Some( x ) => { x.clone() },
_ => { vec![] },
};
let best = actions_percentage.iter().
fold( (Action::NONE, f64::MIN), |accum, x| if x.1 > accum.1 { x.clone() } else { accum } );
let a_text = {
if (j,i) == game._end {
'🏁'
} else if (j,i) == game._start {
'🚶'
} else {
match best {
( Action::UP, _ ) => { '↑' },
( Action::DOWN , _) => { '↓' },
( Action::LEFT, _ ) => { '←' },
( Action::RIGHT, _ ) => { '→' },
_ => { ' ' },
}
}
};
v.push( a_text );
}
println!( "{:?}", v );
}
println!( "linear normalized optimal policy value" );
for i in (0..game._dim.1).rev() {
let mut v = vec![];
for j in 0..game._dim.0 {
let actions_percentage = match policy_normalized.get( &State( j, i ) ) {
Some( x ) => { x.clone() },
_ => { vec![] },
};
let best = actions_percentage.iter().
fold( (Action::NONE, f64::MIN), |accum, x| if x.1 > accum.1 { x.clone() } else { accum } );
let val = match best {
( Action::NONE, _ ) => { 0. },
( _, x ) => { x },
};
v.push( val );
}
println!( "{:?}", v );
}
println!( "policy expectation value map" );
for i in (0..game._dim.1).rev() {
for j in 0..game._dim.0 {
let a = expectation.get( &State( j, i ) ).unwrap_or( & 0. );
print!( "{:.5} ", a );
}
print!( "\n" );
}
}