func NewOracle(env Env) (this *Oracle) {
this = new(Oracle)
this.Env = env
this.Task, _ = rlglue.ParseTaskSpec(this.Env.EnvInit())
this.rehash()
return
}
func (this *OptAgent) AgentInit(taskString string) {
this.task, _ = rlglue.ParseTaskSpec(taskString)
this.Cfg.NumSystems = len(this.task.Obs.Ints)
this.mdp = NewSysMDP(this.Cfg)
this.qt = discrete.NewQTable(this.task.Obs.Ints.Count(), this.task.Act.Ints.Count())
vi.ValueIteration(this.qt, this.mdp, 0.1)
}
func (ra *BebAgent) AgentInit(taskString string) {
ra.task, _ = rlglue.ParseTaskSpec(taskString)
if ra.task.DiscountFactor == 1 {
ra.task.DiscountFactor = 0.99
}
ra.rmdp = NewBebMDP(ra.task, ra.Cfg)
ra.qt = discrete.NewQTable(ra.task.Obs.Ints.Count(), ra.task.Act.Ints.Count())
ra.Cfg.RFoo = ra.GetRFoo(ra.task)
ra.rmdp.RFoo = ra.Cfg.RFoo
}
func (this *Environment) EnvInit() (taskString string) {
fstr := "VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR %f OBSERVATIONS INTS (%d 0 1) ACTIONS INTS (0 %d) REWARDS (-1.0 1.0)"
taskString = fmt.Sprintf(fstr, this.cfg.DiscountFactor, this.cfg.NumSystems, this.cfg.NumSystems)
this.task, _ = rlglue.ParseTaskSpec(taskString)
this.status = make([]bool, this.cfg.NumSystems)
for i := range this.status {
this.status[i] = stat.NextBernoulli(this.cfg.StartBoot) == 1
}
return
}
func TestDepMatch(t *testing.T) {
defer nicetrace.Print()
stat.Seed(seed)
tstr := "VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR 1 OBSERVATIONS INTS (0 4) (-1 1) (0 2) ACTIONS INTS (0 1) REWARDS (0 1.0)"
task, _ := rlglue.ParseTaskSpec(tstr)
stateRanges := task.Obs.Ints
actionRanges := task.Act.Ints
cfg := ConfigDefault()
cfg.Alpha = alpha
cfg.M = M
genDLs := []*DepLearner{NewDepLearner(0, cfg, stateRanges, actionRanges), NewDepLearner(1, cfg, stateRanges, actionRanges), NewDepLearner(2, cfg, stateRanges, actionRanges)}
genDLs[0].SetParents(ParentSet(0).Insert(0, 1))
genDLs[1].SetParents(ParentSet(0).Insert(0, 2))
genDLs[2].SetParents(ParentSet(0).Insert(2))
patDLs := []*DepLearner{NewDepLearner(0, cfg, stateRanges, actionRanges), NewDepLearner(1, cfg, stateRanges, actionRanges), NewDepLearner(2, cfg, stateRanges, actionRanges)}
numStates := stateRanges.Count()
numActions := actionRanges.Count()
RS := stat.Range(int64(numStates))
RA := stat.Range(int64(numActions))
startTime := time.Nanoseconds()
lastWrongStep := make([]int, len(genDLs))
for i := 0; i < steps; i++ {
s := uint64(RS())
a := uint64(RA())
nv := make([]int32, len(genDLs))
for child := 0; child < len(nv); child++ {
nv[child] = genDLs[child].Next(s, a)
}
for child := 0; child < len(nv); child++ {
genDLs[child] = genDLs[child].Update(s, a, nv[child])
}
for child := 0; child < len(nv); child++ {
patDLs[child] = patDLs[child].Update(s, a, nv[child])
}
if i%1 == 0 {
for child := 0; child < len(nv); child++ {
patDLs[child].ConsiderRandomFlip()
if genDLs[child].parents != patDLs[child].parents {
lastWrongStep[child] = i
}
}
}
}
fmt.Println(lastWrongStep)
endTime := time.Nanoseconds()
duration := endTime - startTime
if true {
fmt.Printf("Ran in %fms\n", float64(duration)/1e6)
}
for child := range genDLs {
if genDLs[child].parents != patDLs[child].parents {
t.Error(fmt.Sprintf("%d: %v != %v", child, genDLs[child].parents.Slice(), patDLs[child].parents.Slice()))
}
}
}
func TestBeliefMatch(t *testing.T) {
defer nicetrace.Print()
stat.Seed(seed)
tstr := "VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR 1 OBSERVATIONS INTS (0 4) (-1 1) (0 2) ACTIONS INTS (0 1) REWARDS (0 1.0)"
task, _ := rlglue.ParseTaskSpec(tstr)
cfg := ConfigDefault()
cfg.Alpha = alpha
cfg.M = M
beliefG := NewBelief(cfg, task)
beliefG.learners[0].SetParents(ParentSet(0).Insert(0, 1))
beliefG.learners[1].SetParents(ParentSet(0).Insert(0, 2))
beliefG.learners[2].SetParents(ParentSet(0).Insert(2))
beliefP := NewBelief(cfg, task)
numStates := task.Obs.Ints.Count()
numActions := task.Act.Ints.Count()
RS := stat.Range(int64(numStates))
RA := stat.Range(int64(numActions))
startTime := time.Nanoseconds()
lastWrongStep := make([]int, len(task.Obs.Ints))
for i := 0; i < steps; i++ {
s := uint64(RS())
a := uint64(RA())
n := beliefG.Next(s, a)
beliefG = beliefG.Update(s, a, n).(*Belief)
beliefP = beliefP.Update(s, a, n).(*Belief)
if i%1 == 0 {
beliefP.ConsiderRandomFlipAll()
for child, learner := range beliefP.learners {
if beliefG.learners[child].parents != learner.parents {
lastWrongStep[child] = i
}
}
}
}
fmt.Println(lastWrongStep)
endTime := time.Nanoseconds()
duration := endTime - startTime
if true {
fmt.Printf("Ran in %fms\n", float64(duration)/1e6)
}
for child := range beliefP.learners {
if beliefG.learners[child].parents != beliefP.learners[child].parents {
t.Error(fmt.Sprintf("%d: %v != %v", child, beliefG.learners[child].parents.Slice(), beliefP.learners[child].parents.Slice()))
}
}
}
func NewSysMDP(Cfg Config) (this *SysMDP) {
this = &SysMDP{Cfg: Cfg}
fstr := "VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR %f OBSERVATIONS INTS (%d 0 1) ACTIONS INTS (0 %d) REWARDS (-1.0 1.0)"
taskString := fmt.Sprintf(fstr, this.Cfg.DiscountFactor, this.Cfg.NumSystems, this.Cfg.NumSystems)
this.Task, _ = rlglue.ParseTaskSpec(taskString)
this.maxStates = this.Task.Obs.Ints.Count()
this.maxActions = this.Task.Act.Ints.Count()
this.t = make([][]float64, this.maxStates*this.maxActions)
this.r = make([]float64, this.maxStates*this.maxActions)
for s := range this.S64() {
for a := range this.A64() {
k := s.Hashcode() + a.Hashcode()*this.maxStates
this.t[k] = make([]float64, this.maxStates)
this.r[k] = this.computeR(s, a)
for n := range this.S64() {
this.t[k][n] = this.computeT(s, a, n)
}
}
}
return
}
func NewMDP(cfg Config) (this *MDP) {
this = new(MDP)
this.Cfg = cfg
fstr := "VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR %f OBSERVATIONS INTS (6 0 1) ACTIONS INTS (0 3) REWARDS (%f %f)"
rmin := 0.0
rmax := 1.0
taskString := fmt.Sprintf(fstr, this.Cfg.DiscountFactor, rmin, rmax)
this.Task, _ = rlglue.ParseTaskSpec(taskString)
this.maxStates = this.Task.Obs.Ints.Count()
this.maxActions = this.Task.Act.Ints.Count()
this.svs = make([][]int32, this.maxStates)
for s := uint64(0); s < this.maxStates; s++ {
this.svs[s] = this.Task.Obs.Ints.Values(s)
}
this.t = make([][]float64, this.maxStates*this.maxActions)
this.r = make([]float64, this.maxStates*this.maxActions)
for s := range this.S64() {
for a := range this.A64() {
k := s.Hashcode() + this.maxStates*a.Hashcode()
this.r[k] = this.computeR(s, a)
this.t[k] = make([]float64, this.maxStates)
sum := 0.0
for n := range this.S64() {
this.t[k][n] = this.computeT(s, a, n)
sum += this.t[k][n]
}
if sum > 1 {
fmt.Fprintf(os.Stderr, "%v %v\n", this.svs[s], a)
for n := range this.S64() {
if this.t[k][n] != 0 {
fmt.Fprintf(os.Stderr, "\t%v : %v\n", this.svs[n], this.t[k][n])
}
}
}
}
}
return
}
func (this *ROARAgent) AgentInit(taskString string) {
this.task, _ = rlglue.ParseTaskSpec(taskString)
this.numFeatures = len(this.task.Obs.Doubles)
}
func (ra *RmaxAgent) AgentInit(taskString string) {
ra.task, _ = rlglue.ParseTaskSpec(taskString)
ra.rmdp = NewRmaxMDP(ra.task, ra.Cfg.M)
ra.qt = discrete.NewQTable(ra.task.Obs.Ints.Count(), ra.task.Act.Ints.Count())
}
func (this *BFS3Agent) AgentInit(taskString string) {
this.task, _ = rlglue.ParseTaskSpec(taskString)
this.belief = this.prior(this.task)
this.ResetPlanner()
}
package wumpus
import (
"fmt"
"github.com/skelterjohn/rlbayes"
"go-glue.googlecode.com/hg/rlglue"
"go-glue.googlecode.com/hg/rltools/discrete"
)
var taskstr = fmt.Sprintf("VERSION RL-Glue-3.0 PROBLEMTYPE episodic DISCOUNTFACTOR %f OBSERVATIONS INTS (2 0 3) (0 3) (16 0 %d) ACTIONS INTS (0 3) REWARDS (%f, 1)", Gamma, len(SquareMapToValue), -.01/(1-Gamma))
var task, _ = rlglue.ParseTaskSpec(taskstr)
type Belief struct {
Hunter
term bool
hash uint64
}
func NewBelief(mb MapBelief) (this *Belief) {
this = new(Belief)
this.x = 0
this.y = 0
this.dir = 1
this.belief = mb
return
}
func (this *Belief) Hashcode() uint64 {
return this.hash
}
func (this *Belief) LessThan(oi interface{}) bool {
return this.hash < oi.(*Belief).hash
func (ra *RmaxFSSSAgent) AgentInit(taskString string) {
ra.task, _ = rlglue.ParseTaskSpec(taskString)
ra.rmdp = rmax.NewRmaxMDP(ra.task, ra.Cfg.M)
}
