Initial working alphazero agent.

.github/workflows/main.yml

@@ -31,3 +31,5 @@ jobs:
       - name: pytest
         run: |
           cd python && pytest
+        env:
+          TEST_WANDB: true

Makefile

@@ -1,7 +1,7 @@
 PREFIX := github.com/nelhage/taktician
 
-PROTOS := $(wildcard proto/*.proto)
-PROTONAMES := $(basename $(notdir $(PROTOS)))
+PROTOS := $(wildcard proto/tak/proto/*.proto)
+PROTONAMES := $(foreach proto,$(PROTOS), $(basename $(notdir $(proto))))
 GOPROTOSRC := $(foreach proto,$(PROTONAMES),pb/$(proto).pb.go)
 PYPROTOSRC := $(foreach proto,$(PROTONAMES),python/tak/proto/$(proto)_pb2.py)
 GENFILES := ai/feature_string.go $(GOPROTOSRC) $(PYPROTOSRC)
@@ -13,9 +13,15 @@ ai/feature_string.go: ai/evaluate.go
 protoc: $(GOPROTOSRC) $(PYPROTOSRC)
 
 $(GOPROTOSRC) $(PYPROTOSRC): $(PROTOS)
-	protoc -I proto/ \
-	       --python_out=python/tak/proto/ --go_out=pb \
-	       proto/*.proto
+	python -m grpc_tools.protoc\
+	       -I proto/ \
+	       --python_out=python/ \
+	       --grpc_python_out=python/ \
+	       --go_out=. \
+	       --go_opt="module=$(PREFIX)" \
+	       --go-grpc_out=. \
+	       --go-grpc_opt="module=$(PREFIX)" \
+	       proto/tak/proto/*.proto
 
 build: $(GENFILES)
 	go build $(PREFIX)/...

ai/evaluate.go

@@ -341,7 +341,7 @@ func scoreGroups(c *bitboard.Constants, gs []uint64, ws *Weights, other uint64)
 	return sc
 }
 
-func countThreats(c *bitboard.Constants, p *tak.Position) (wp, wt, bp, bt int) {
+func CountThreats(c *bitboard.Constants, p *tak.Position) (wp, wt, bp, bt int) {
 	analysis := p.Analysis()
 	empty := c.Mask &^ (p.White | p.Black)
 
@@ -416,7 +416,7 @@ func scoreThreats(c *bitboard.Constants, ws *Weights, p *tak.Position) int64 {
 		return 0
 	}
 
-	wp, wt, bp, bt := countThreats(c, p)
+	wp, wt, bp, bt := CountThreats(c, p)
 
 	if wp+wt > 0 && p.ToMove() == tak.White {
 		return ForcedWin
@@ -543,7 +543,7 @@ func ExplainScore(m *MinimaxAI, out io.Writer, p *tak.Position) {
 
 	fmt.Fprintf(tw, "liberties\t%d\t%d\n", wl, bl)
 
-	wp, wt, bp, bt := countThreats(&m.c, p)
+	wp, wt, bp, bt := CountThreats(&m.c, p)
 	fmt.Fprintf(tw, "potential\t%d\t%d\n", wp, bp)
 	fmt.Fprintf(tw, "threat\t%d\t%d\n", wt, bt)
 

ai/mcts/debug.go

@@ -27,6 +27,10 @@ func (mc *MonteCarloAI) dumpTreeNode(f io.Writer, t *tree) {
 	parent := 1
 	if t.parent != nil {
 		parent = t.parent.simulations
+		if t.parent.proven != 0 && t.proven == 0 {
+			return
+		}
+
 	}
 	label := fmt.Sprintf("n=%d p=%d v=%.0f+%.0f",
 		t.simulations,
@@ -36,17 +40,20 @@ func (mc *MonteCarloAI) dumpTreeNode(f io.Writer, t *tree) {
 
 	fmt.Fprintf(f, `  n%p [label="%s"]`, t, label)
 	fmt.Fprintln(f)
-	if t.children == nil || t.simulations < mc.cfg.InitialVisits {
+	if t.children == nil {
 		return
 	}
 
 	for _, c := range t.children {
-		if c.simulations < mc.cfg.InitialVisits {
+		if t.proven > 0 && c.proven >= 0 {
 			continue
 		}
 		fmt.Fprintf(f, `  n%p -> n%p [label="%s"]`,
 			t, c, ptn.FormatMove(c.move))
 		fmt.Fprintln(f)
 		mc.dumpTreeNode(f, c)
+		if c.proven < 0 {
+			break
+		}
 	}
 }

ai/mcts/mcts.go

@@ -22,7 +22,6 @@ type MCTSConfig struct {
 	C     float64
 	Seed  int64
 
-	InitialVisits int
 	MMDepth       int
 	MaxRollout    int
 	EvalThreshold int64
@@ -154,21 +153,6 @@ func (ai *MonteCarloAI) GetMove(ctx context.Context, p *tak.Position) tak.Move {
 		ai.dumpTree(tree)
 	}
 
-	if tree.proven != 0 {
-		if len(tree.children) == 0 {
-			return ai.mm.GetMove(ctx, p)
-		}
-		best := tree.children[0]
-		for _, c := range tree.children {
-			if c.proven < best.proven {
-				best = c
-			}
-		}
-		if ai.cfg.Debug > 1 {
-			log.Printf("proven m=%s v=%d", ptn.FormatMove(best.move), -best.proven)
-		}
-		return best.move
-	}
 	best := tree.children[0]
 	i := 0
 	sort.Sort(bySims(tree.children))
@@ -194,6 +178,21 @@ func (ai *MonteCarloAI) GetMove(ctx context.Context, p *tak.Position) tak.Move {
 			}
 		}
 	}
+	if tree.proven != 0 {
+		if len(tree.children) == 0 {
+			return ai.mm.GetMove(ctx, p)
+		}
+		best := tree.children[0]
+		for _, c := range tree.children {
+			if c.proven < best.proven {
+				best = c
+			}
+		}
+		if ai.cfg.Debug > 1 {
+			log.Printf("proven m=%s v=%d", ptn.FormatMove(best.move), -best.proven)
+		}
+		return best.move
+	}
 	if ai.cfg.Debug > 0 {
 		log.Printf("[mcts] evaluated simulations=%d value=%d proven=%d", tree.simulations, tree.value, tree.proven)
 	}
@@ -214,14 +213,16 @@ func (mc *MonteCarloAI) printdbg(t *tree) {
 }
 
 func (mc *MonteCarloAI) populate(ctx context.Context, t *tree) {
-	_, v, _ := mc.mm.Analyze(ctx, t.position)
-	if v > ai.WinThreshold {
-		t.proven = 1
-		return
-	} else if v < -ai.WinThreshold {
-		t.proven = -1
-		return
-	}
+	/*
+		_, v, _ := mc.mm.Analyze(ctx, t.position)
+		if v > ai.WinThreshold {
+			t.proven = 1
+			return
+		} else if v < -ai.WinThreshold {
+			t.proven = -1
+			return
+		}
+	*/
 
 	moves := t.position.AllMoves(nil)
 	t.children = make([]*tree, 0, len(moves))
@@ -354,9 +355,6 @@ func NewMonteCarlo(cfg MCTSConfig) *MonteCarloAI {
 	if mc.cfg.Seed == 0 {
 		mc.cfg.Seed = time.Now().Unix()
 	}
-	if mc.cfg.InitialVisits == 0 {
-		mc.cfg.InitialVisits = 3
-	}
 	if mc.cfg.MMDepth == 0 {
 		mc.cfg.MMDepth = 3
 	}

ai/random.go

@@ -13,7 +13,8 @@ type RandomAI struct {
 }
 
 func (r *RandomAI) GetMove(ctx context.Context, p *tak.Position) tak.Move {
-	moves := p.AllMoves(nil)
+	var buffer [100]tak.Move
+	moves := p.AllMoves(buffer[:0])
 	i := r.r.Int31n(int32(len(moves)))
 	return moves[i]
 }

bin/sync

@@ -0,0 +1,24 @@
+#!/bin/bash
+set -eu
+
+root=$(dirname "$0")/../
+cd "$root"
+target=$1
+path=${2-'~/code/taktician'}
+
+
+exec rsync -Pax \
+     --exclude='__pycache__' \
+     --exclude='/python/notebooks/' \
+     --exclude='/python/build/' \
+     --exclude='/python/dist/' \
+     --exclude='/python/.pytest_cache/' \
+     --exclude='/python/*.so' \
+     --exclude='/.git' \
+     --exclude='/.direnv' \
+     --exclude='*.test' \
+     --exclude='/data' \
+     --exclude='wandb' \
+     --exclude='/.envrc' \
+     --exclude='python/tak.egg-info' \
+     . "$target:$path"

cmd/internal/analyze/analysis.go

@@ -110,7 +110,7 @@ func (a *pnAnalysis) Analyze(ctx context.Context, p *tak.Position) {
 		cli.RenderBoard(nil, os.Stdout, p)
 	}
 
-	out := prover.Prove(ctx, p)
+	out, stats := prover.Prove(ctx, p)
 	var result string
 	switch out.Result {
 	case prove.EvalTrue:
@@ -131,11 +131,11 @@ func (a *pnAnalysis) Analyze(ctx context.Context, p *tak.Position) {
 		result,
 		move,
 		out.Duration,
-		out.Stats.Nodes,
+		stats.Nodes,
 		out.Proof,
 		out.Disproof,
 		out.Depth,
-		out.Stats.MaxDepth,
+		stats.MaxDepth,
 	)
 
 	if a.cmd.dumpTree != "" {
@@ -151,3 +151,61 @@ func (a *pnAnalysis) Analyze(ctx context.Context, p *tak.Position) {
 		out.Close()
 	}
 }
+
+type dfpnAnalysis struct {
+	cmd *Command
+}
+
+func (a *dfpnAnalysis) Analyze(ctx context.Context, p *tak.Position) {
+	var attacker tak.Color
+	if a.cmd.attacker == "white" {
+		attacker = tak.White
+	} else if a.cmd.attacker == "black" {
+		attacker = tak.Black
+	} else if a.cmd.attacker == "" {
+		attacker = tak.NoColor
+	} else {
+		log.Fatalf("Cannot parse attacker: %q", a.cmd.attacker)
+	}
+	prover := prove.NewDFPN(&prove.DFPNConfig{
+		Debug:    a.cmd.mmopt.Debug,
+		TableMem: a.cmd.mmopt.TableMem,
+		Attacker: attacker,
+	})
+
+	if !a.cmd.quiet {
+		cli.RenderBoard(nil, os.Stdout, p)
+	}
+
+	out, stats := prover.Prove(p)
+	var result string
+	switch out.Result {
+	case prove.EvalTrue:
+		result = "WIN"
+	case prove.EvalFalse:
+		result = "DRAW|LOSE"
+	case prove.EvalUnknown:
+		result = "UNKNOWN"
+	}
+	fmt.Printf("PN search analysis:\n")
+	var move string
+	if out.Move.Type != 0 {
+		move = ptn.FormatMove(out.Move)
+	} else {
+		move = "(none)"
+	}
+	fmt.Printf(" value=%s move=%s duration=%s\n",
+		result,
+		move,
+		out.Duration,
+	)
+	fmt.Printf(" work=%d terminal=%d solved=%d repetition=%d hit=%d/%d (%0.2f%%)\n",
+		stats.Work,
+		stats.Terminal,
+		stats.Solved,
+		stats.Repetition,
+		stats.Hits,
+		stats.Hits+stats.Miss,
+		100*float64(stats.Hits)/float64(stats.Hits+stats.Miss),
+	)
+}