func TestEnvTraj(t *testing.T) {
if false {
return
}
stat.TimeSeed()
defer nicetrace.Print()
env := New()
obsi := env.EnvStart()
indexi := task.Obs.Ints.Index(obsi.Ints())
b := NewBelief(make(MapBelief, 16))
b.Teleport(indexi)
fmt.Printf("%v\n", b.Hunter)
do := func(what int32) bool {
action := rlglue.NewAction([]int32{what}, []float64{}, []byte{})
obs, r, t := env.EnvStep(action)
fmt.Println(what, r)
if t {
fmt.Println("done")
return false
}
index := task.Obs.Ints.Index(obs.Ints())
bs := b.Update(uint64(what), index, r)
b = bs.(*Belief)
fmt.Printf("%v\n", b.Hunter)
return true
}
guess := func() (what int32) {
fmt.Scanf("%d", &what)
return
}
for do(guess()) {
}
}
func (ra *RmaxFSSSAgent) AgentStart(obs rlglue.Observation) (act rlglue.Action) {
ra.stepsWithPlanner = 0
ra.lastState = discrete.State(ra.task.Obs.Ints.Index(obs.Ints()))
ra.Plan()
act = rlglue.NewAction(ra.task.Act.Ints.Values(ra.GetAction().Hashcode()), []float64{}, []byte{})
ra.lastAction = discrete.Action(ra.task.Act.Ints.Index(act.Ints()))
return
}
func (this *Agent) AgentStart(obs rlglue.Observation) (act rlglue.Action) {
this.stepsWithPlanner = 0
this.lastState = discrete.State(this.mdp.GetTask().Obs.Ints.Index(obs.Ints()))
this.Plan()
act = rlglue.NewAction(this.mdp.GetTask().Act.Ints.Values(this.GetAction()), []float64{}, []byte{})
this.lastAction = discrete.Action(this.mdp.GetTask().Act.Ints.Index(act.Ints()))
return
}
func (this *Environment) Next(action discrete.Action) (o discrete.Oracle, r float64) {
act := rlglue.NewAction([]int32{int32(action)}, []float64{}, []byte{})
next := new(Environment)
*next = *this
next.status = append([]bool{}, this.status...)
_, r, _ = next.EnvStep(act)
o = next
return
}
func (ra *BebAgent) AgentStep(reward float64, obs rlglue.Observation) (act rlglue.Action) {
nextState := discrete.State(ra.task.Obs.Ints.Index(obs.Ints()))
learned := ra.rmdp.Observe(ra.lastState, ra.lastAction, nextState, reward)
if learned {
vi.ValueIteration(ra.qt, ra.rmdp, ra.Cfg.Epsilon)
}
ra.lastState = nextState
act = rlglue.NewAction(ra.task.Act.Ints.Values(ra.qt.Pi(ra.lastState).Hashcode()), []float64{}, []byte{})
ra.lastAction = discrete.Action(ra.task.Act.Ints.Index(act.Ints()))
return
}
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
}
func (ra *RmaxFSSSAgent) AgentStep(reward float64, obs rlglue.Observation) (act rlglue.Action) {
ra.stepsWithPlanner++
nextState := discrete.State(ra.task.Obs.Ints.Index(obs.Ints()))
learned := ra.rmdp.Observe(ra.lastState, ra.lastAction, nextState, reward)
if learned {
ra.Forget()
}
ra.lastState = nextState
ra.Plan()
act = rlglue.NewAction(ra.task.Act.Ints.Values(ra.GetAction().Hashcode()), []float64{}, []byte{})
ra.lastAction = discrete.Action(ra.task.Act.Ints.Index(act.Ints()))
return
}
func (ra *BebAgent) AgentStart(obs rlglue.Observation) (act rlglue.Action) {
ra.lastState = discrete.State(ra.task.Obs.Ints.Index(obs.Ints()))
act = rlglue.NewAction(ra.task.Act.Ints.Values(ra.qt.Pi(ra.lastState).Hashcode()), []float64{}, []byte{})
ra.lastAction = discrete.Action(ra.task.Act.Ints.Index(act.Ints()))
return
}
func (this *ROARAgent) GetAction() (act rlglue.Action) {
index := uint64(rand.Int63n(int64(this.task.Obs.Ints.Count())))
act = rlglue.NewAction(this.task.Obs.Ints.Values(index), []float64{}, []byte{})
this.LastAct = act
return
}
func (this *BFS3Agent) getIndexAction(index discrete.Action) (act rlglue.Action) {
return rlglue.NewAction(this.task.Act.Ints.Values(index.Hashcode()), []float64{}, []byte{})
}
func (this *OptAgent) AgentStep(reward float64, obs rlglue.Observation) (act rlglue.Action) {
s := discrete.State(this.task.Obs.Ints.Index(obs.Ints()))
a := this.qt.Pi(s)
act = rlglue.NewAction([]int32{int32(a)}, []float64{}, []byte{})
return
}
