Ejemplo n.º 1
0
func (this *CRPReward) Update(s discrete.State, a discrete.Action, r float64) (next RewardBelief) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		return this
	}
	ndr := new(CRPReward)
	*ndr = *this
	ndr.Known = make([]bool, len(this.Known))
	copy(ndr.Known, this.Known)
	ndr.R = make([]float64, len(this.R))
	copy(ndr.R, this.R)
	ndr.Known[index] = true
	ndr.R[index] = r
	ndr.countKnown++

	ndr.SeenRewards = append([]float64{r}, this.SeenRewards...)
	ndr.Counts = append([]uint64{1}, this.Counts...)
	var seen bool
	for i, sr := range this.SeenRewards {
		if i != 0 && sr == r {
			seen = true
			ndr.Counts[i]++
			break
		}
	}
	if seen {
		ndr.SeenRewards = ndr.SeenRewards[1:len(ndr.SeenRewards)]
		ndr.Counts = ndr.Counts[1:len(ndr.Counts)]
	}

	ndr.Total++

	next = ndr
	return
}
Ejemplo n.º 2
0
func (this *CRPReward) Next(s discrete.State, a discrete.Action) (r float64) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		r = this.R[index]
		return
	}

	if this.chooser == nil {
		if len(this.Counts) == 0 {
			this.chooser = func() int64 { return 0 }
		} else {
			normalizer := 1.0 / (float64(this.Total) + this.Alpha)
			weights := make([]float64, len(this.Counts))
			for i := range weights {
				weights[i] = float64(this.Counts[i]) * normalizer
			}
			this.chooser = stat.Choice(weights)
		}
	}

	which := int(this.chooser())
	if which == len(this.SeenRewards) {
		r = this.BaseSampler()
	} else {
		r = this.SeenRewards[which]
	}

	return
}
Ejemplo n.º 3
0
func (this *FDMTransition) Update(s discrete.State, a discrete.Action, n discrete.State) (next TransitionBelief) {
	o := this.bg.NextToOutcome(s, n)
	k := s.Hashcode() + a.Hashcode()*this.bg.NumStates
	dsa := this.sas[k]
	if dsa == nil {
		dsa = NewDirSA(this.bg.Alpha)
		this.sas[k] = dsa
	}

	if this.bg.ForgetThreshold != 0 && dsa.visits >= this.bg.ForgetThreshold {
		next = this
		return
	}

	nextFDM := new(FDMTransition)
	nextFDM.bg = this.bg
	nextFDM.sas = make([]*DirSA, len(this.sas))
	copy(nextFDM.sas, this.sas)
	nextFDM.sas[k] = dsa.Update(o)
	if nextFDM.sas[k].visits == this.bg.ForgetThreshold {
		nextFDM.sas[k].ForgetPrior(this.bg.Alpha)
		//fmt.Printf("%v\n", nextFDM.sas[k])
	}
	nextFDM.hash = this.hash - this.sas[k].Hashcode() + nextFDM.sas[k].Hashcode()
	next = nextFDM

	return
}
Ejemplo n.º 4
0
func (this *DeterministicReward) Next(s discrete.State, a discrete.Action) (r float64) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		r = this.R[index]
		return
	}
	r = this.BaseSampler()
	return
}
Ejemplo n.º 5
0
func (this *FObjTransition) Next(s discrete.State, a discrete.Action) (n discrete.State) {
	avalues := this.bg.Task.Act.Ints.Values(a.Hashcode())
	which, act := avalues[0], avalues[1]
	sobjs := this.bg.GetObjs(s)
	nobjs := append([]discrete.State{}, sobjs...)
	nobjs[which] = this.ObjFDM.Next(sobjs[which], discrete.Action(act))
	n = this.bg.GetState(nobjs)
	return
}
Ejemplo n.º 6
0
func (this *Posterior) UpdatePosterior(s discrete.State, a discrete.Action, o discrete.State) (next *Posterior) {
	next = new(Posterior)
	*next = *this
	next.stateData = append([]SAHist{}, this.stateData...)
	next.clusterData = append([]SAHist{}, this.clusterData...)
	next.C = this.C.Copy()
	k := s.Hashcode()*this.bg.NumActions + a.Hashcode()
	next.stateData[k] = next.stateData[k].Incr(this.bg.NumOutcomes, o)
	return
}
Ejemplo n.º 7
0
func (this *FDMTransition) Next(s discrete.State, a discrete.Action) (n discrete.State) {
	k := s.Hashcode() + a.Hashcode()*this.bg.NumStates
	dsa := this.sas[k]
	if dsa == nil {
		dsa = NewDirSA(this.bg.Alpha)
		this.sas[k] = dsa
	}
	n = this.bg.OutcomeToNext(s, dsa.Next())
	return
}
Ejemplo n.º 8
0
func (this *Oracle) Next(action discrete.Action) (o discrete.Oracle, r float64) {
	avalues := this.Task.Act.Ints.Values(action.Hashcode())
	act := rlglue.NewAction(avalues, []float64{}, []byte{})
	next := new(Oracle)
	*next = *this
	next.Cans = append([]Can{}, this.Cans...)
	_, r, next.isTerminal = next.Env.EnvStep(act)
	next.rehash()
	o = next
	return
}
Ejemplo n.º 9
0
func (this *FObjTransition) Update(s discrete.State, a discrete.Action, n discrete.State) (next TransitionBelief) {
	nt := new(FObjTransition)
	*nt = *this
	avalues := this.bg.Task.Act.Ints.Values(a.Hashcode())
	which, act := avalues[0], avalues[1]
	sobjs := this.bg.GetObjs(s)
	nobjs := this.bg.GetObjs(n)
	nt.ObjFDM = this.ObjFDM.Update(sobjs[which], discrete.Action(act), nobjs[which]).(*FDMTransition)
	next = nt
	return
}
Ejemplo n.º 10
0
func (this *BetaTerminal) Next(s discrete.State, a discrete.Action) (t bool) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		t = this.Term[index]
		return
	}
	prob := this.Alpha / (this.Alpha + this.Beta)
	if stat.NextUniform() < prob {
		t = true
	}
	return
}
Ejemplo n.º 11
0
func (this *CountKnown) Update(s discrete.State, a discrete.Action) (next KnownBelief) {
	nk := new(CountKnown)
	nk.numStates = this.numStates
	nk.visits = make([]int, len(this.visits))
	copy(nk.visits, this.visits)
	nk.threshold = this.threshold

	k := s.Hashcode() + nk.numStates*a.Hashcode()

	nk.visits[k]++
	next = nk
	return
}
Ejemplo n.º 12
0
func (this *RmaxReward) Update(s discrete.State, a discrete.Action, r float64) (next RewardBelief) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.R[index] == r {
		return this
	}
	nrr := new(RmaxReward)
	*nrr = *this
	nrr.R = make([]float64, len(this.R))
	copy(nrr.R, this.R)
	nrr.R[index] = r
	nrr.countKnown++
	return nrr
}
Ejemplo n.º 13
0
func (this *Posterior) Next(s discrete.State, a discrete.Action) (n discrete.State) {
	c := uint64(this.C.Get(int(s)))
	ck := c*this.bg.NumActions + a.Hashcode()
	hist := this.clusterData[ck]
	fhist := append([]float64{}, this.bg.Beta...)
	total := 0.0
	for i, c := range hist {
		fhist[i] += float64(c)
		total += fhist[i]
	}
	for i := range fhist {
		fhist[i] /= total
	}
	o := discrete.State(stat.NextChoice(fhist))
	n = this.bg.OutcomeToNext(s, o)
	return
}
Ejemplo n.º 14
0
func (this *DeterministicReward) Update(s discrete.State, a discrete.Action, r float64) (next RewardBelief) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		return this
	}
	ndr := new(DeterministicReward)
	*ndr = *this
	ndr.Known = make([]bool, len(this.Known))
	copy(ndr.Known, this.Known)
	ndr.R = make([]float64, len(this.R))
	copy(ndr.R, this.R)
	ndr.Known[index] = true
	ndr.R[index] = r
	ndr.countKnown++
	next = ndr
	return
}
Ejemplo n.º 15
0
func (this *DepLearner) Update(s discrete.State, a discrete.Action, o int32) (next *DepLearner) {
	k := a.Hashcode() + this.bg.numActions*s.Hashcode()
	next = new(DepLearner)
	*next = *this
	oi := this.bg.myRange.Index(o)
	next.history = append([]Histogram{}, this.history...)
	next.history[k] = next.history[k].Incr(oi)
	sv := next.bg.stateValues[s]
	mv := next.parents.CutValues(sv)
	ms := next.cutRanges.Index(mv)
	mk := a.Hashcode() + this.bg.numActions*ms
	next.mappedHistory = append([]Histogram{}, this.mappedHistory...)
	next.mappedLoglihood += next.mappedHistory[mk].LogFactorAlpha(this.bg.cfg.Alpha)
	next.mappedHistory[mk] = next.mappedHistory[mk].Incr(oi)
	next.mappedLoglihood -= next.mappedHistory[mk].LogFactorAlpha(this.bg.cfg.Alpha)
	next.hash += k << oi
	return
}
Ejemplo n.º 16
0
func (this *DepLearner) Next(s discrete.State, a discrete.Action) (o int32) {
	sv := this.bg.stateValues[s]
	mv := this.parents.CutValues(sv)
	ms := this.cutRanges.Index(mv)
	mk := a.Hashcode() + this.bg.numActions*ms
	h := this.mappedHistory[mk]
	lls := make([]float64, len(h))
	usePrior := h.Sum() < this.bg.cfg.M
	for i, c := range h {
		if usePrior {
			lls[i] = math.Log(this.bg.cfg.Alpha + float64(c))
		} else {
			lls[i] = math.Log(float64(c))
		}
	}
	oi := uint64(stat.NextLogChoice(lls))
	o = this.bg.myRange.Value(oi)
	return
}
Ejemplo n.º 17
0
func (this *BetaTerminal) Update(s discrete.State, a discrete.Action, t bool) (next TerminalBelief) {
	index := s.Hashcode() + this.NumStates*a.Hashcode()
	if this.Known[index] {
		next = this
		return
	}
	nbt := new(BetaTerminal)
	*nbt = *this
	nbt.Known = append([]bool{}, this.Known...)
	nbt.Term = append([]bool{}, this.Term...)
	nbt.Known[index] = true
	nbt.Term[index] = t
	if t {
		nbt.Alpha++
	} else {
		nbt.Beta++
	}

	return nbt
}
Ejemplo n.º 18
0
func (this *Belief) Update(s discrete.State, a discrete.Action, n discrete.State) (nextBelief bayes.TransitionBelief) {
	k := a.Hashcode() + s.Hashcode()*this.bg.numActions
	if this.totals[k] >= this.bg.cfg.M {
		nextBelief = this
		return
	}
	nv := this.bg.stateValues[n]
	next := new(Belief)
	*next = *this
	next.hash = 0
	next.learners = append([]*DepLearner{}, this.learners...)
	for child := range this.learners {
		next.learners[child] = next.learners[child].Update(s, a, nv[child])
		next.hash += next.learners[child].Hashcode() << uint(child)
	}
	next.totals = append([]uint64{}, this.totals...)
	next.totals[k]++
	nextBelief = next
	return
}
Ejemplo n.º 19
0
func (this *BFS3Agent) getIndexAction(index discrete.Action) (act rlglue.Action) {
	return rlglue.NewAction(this.task.Act.Ints.Values(index.Hashcode()), []float64{}, []byte{})
}
Ejemplo n.º 20
0
func (this *RmaxReward) Next(s discrete.State, a discrete.Action) (r float64) {
	return this.R[s.Hashcode()+this.NumStates*a.Hashcode()]
}
Ejemplo n.º 21
0
func (this *CountKnown) Known(s discrete.State, a discrete.Action) (known bool) {
	k := s.Hashcode() + this.numStates*a.Hashcode()
	return this.visits[k] >= this.threshold
}
Ejemplo n.º 22
0
func (this *SysMDP) R(s discrete.State, a discrete.Action) float64 {
	k := s.Hashcode() + a.Hashcode()*this.maxStates
	return this.r[k]
}
Ejemplo n.º 23
0
func (this *SysMDP) T(s discrete.State, a discrete.Action, n discrete.State) float64 {
	k := s.Hashcode() + a.Hashcode()*this.maxStates
	return this.t[k][n]
}