init space
Browse files- Dockerfile +3 -3
- LightZero/.gitignore +0 -13
- LightZero/lzero/mcts/ctree/common_lib/cminimax.cpp +68 -0
- LightZero/lzero/mcts/ctree/common_lib/cminimax.h +40 -0
- LightZero/lzero/mcts/ctree/common_lib/utils.cpp +27 -0
- LightZero/lzero/mcts/ctree/ctree_alphazero/mcts_alphazero.cpp +348 -0
- LightZero/lzero/mcts/ctree/ctree_alphazero/node_alphazero.cpp +22 -0
- LightZero/lzero/mcts/ctree/ctree_alphazero/node_alphazero.h +85 -0
Dockerfile
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
FROM
|
2 |
|
3 |
ENV DEBIAN_FRONTEND=noninteractive
|
4 |
ENV LANG en_US.UTF-8
|
@@ -7,9 +7,9 @@ ENV LC_ALL en_US.UTF-8
|
|
7 |
|
8 |
RUN apt update -y \
|
9 |
&& apt install libgl1-mesa-glx libglib2.0-0 libsm6 libxext6 libxrender-dev swig curl git vim gcc \g++ make wget locales dnsutils zip unzip cmake nginx -y \
|
10 |
-
&& curl -fsSL https://deb.nodesource.com/
|
11 |
&& apt-get install -y nodejs \
|
12 |
-
&& npm install -g npm@
|
13 |
&& npm install -g create-react-app \
|
14 |
&& npm install typescript -g \
|
15 |
&& npm install -g vite \
|
|
|
1 |
+
FROM pytorch/pytorch:2.1.2-cuda12.1-cudnn8-runtime as base
|
2 |
|
3 |
ENV DEBIAN_FRONTEND=noninteractive
|
4 |
ENV LANG en_US.UTF-8
|
|
|
7 |
|
8 |
RUN apt update -y \
|
9 |
&& apt install libgl1-mesa-glx libglib2.0-0 libsm6 libxext6 libxrender-dev swig curl git vim gcc \g++ make wget locales dnsutils zip unzip cmake nginx -y \
|
10 |
+
&& curl -fsSL https://deb.nodesource.com/setup_20.x | bash - \
|
11 |
&& apt-get install -y nodejs \
|
12 |
+
&& npm install -g npm@10.3.0 \
|
13 |
&& npm install -g create-react-app \
|
14 |
&& npm install typescript -g \
|
15 |
&& npm install -g vite \
|
LightZero/.gitignore
CHANGED
@@ -1432,16 +1432,3 @@ collect_demo_data_config.py
|
|
1432 |
events.*
|
1433 |
/test_*
|
1434 |
# LightZero special key
|
1435 |
-
/zoo/board_games/**/*.c
|
1436 |
-
/zoo/board_games/**/*.cpp
|
1437 |
-
/lzero/mcts/**/*.cpp
|
1438 |
-
/zoo/**/*.c
|
1439 |
-
/lzero/mcts/**/*.so
|
1440 |
-
/lzero/mcts/**/*.h
|
1441 |
-
!/lzero/mcts/**/lib
|
1442 |
-
!/lzero/mcts/**/lib/*.cpp
|
1443 |
-
!/lzero/mcts/**/lib/*.hpp
|
1444 |
-
!/lzero/mcts/**/lib/*.h
|
1445 |
-
**/tb/*
|
1446 |
-
**/mcts/ctree/tests_cpp/*
|
1447 |
-
**/*tmp*
|
|
|
1432 |
events.*
|
1433 |
/test_*
|
1434 |
# LightZero special key
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
LightZero/lzero/mcts/ctree/common_lib/cminimax.cpp
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
// C++11
|
2 |
+
|
3 |
+
#include "cminimax.h"
|
4 |
+
|
5 |
+
namespace tools{
|
6 |
+
|
7 |
+
CMinMaxStats::CMinMaxStats(){
|
8 |
+
this->maximum = FLOAT_MIN;
|
9 |
+
this->minimum = FLOAT_MAX;
|
10 |
+
this->value_delta_max = 0.;
|
11 |
+
}
|
12 |
+
|
13 |
+
CMinMaxStats::~CMinMaxStats(){}
|
14 |
+
|
15 |
+
void CMinMaxStats::set_delta(float value_delta_max){
|
16 |
+
this->value_delta_max = value_delta_max;
|
17 |
+
}
|
18 |
+
|
19 |
+
void CMinMaxStats::update(float value){
|
20 |
+
if(value > this->maximum){
|
21 |
+
this->maximum = value;
|
22 |
+
}
|
23 |
+
if(value < this->minimum){
|
24 |
+
this->minimum = value;
|
25 |
+
}
|
26 |
+
}
|
27 |
+
|
28 |
+
void CMinMaxStats::clear(){
|
29 |
+
this->maximum = FLOAT_MIN;
|
30 |
+
this->minimum = FLOAT_MAX;
|
31 |
+
}
|
32 |
+
|
33 |
+
float CMinMaxStats::normalize(float value){
|
34 |
+
float norm_value = value;
|
35 |
+
float delta = this->maximum - this->minimum;
|
36 |
+
if(delta > 0){
|
37 |
+
if(delta < this->value_delta_max){
|
38 |
+
norm_value = (norm_value - this->minimum) / this->value_delta_max;
|
39 |
+
}
|
40 |
+
else{
|
41 |
+
norm_value = (norm_value - this->minimum) / delta;
|
42 |
+
}
|
43 |
+
}
|
44 |
+
return norm_value;
|
45 |
+
}
|
46 |
+
|
47 |
+
//*********************************************************
|
48 |
+
|
49 |
+
CMinMaxStatsList::CMinMaxStatsList(){
|
50 |
+
this->num = 0;
|
51 |
+
}
|
52 |
+
|
53 |
+
CMinMaxStatsList::CMinMaxStatsList(int num){
|
54 |
+
this->num = num;
|
55 |
+
for(int i = 0; i < num; ++i){
|
56 |
+
this->stats_lst.push_back(CMinMaxStats());
|
57 |
+
}
|
58 |
+
}
|
59 |
+
|
60 |
+
CMinMaxStatsList::~CMinMaxStatsList(){}
|
61 |
+
|
62 |
+
void CMinMaxStatsList::set_delta(float value_delta_max){
|
63 |
+
for(int i = 0; i < this->num; ++i){
|
64 |
+
this->stats_lst[i].set_delta(value_delta_max);
|
65 |
+
}
|
66 |
+
}
|
67 |
+
|
68 |
+
}
|
LightZero/lzero/mcts/ctree/common_lib/cminimax.h
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
// C++11
|
2 |
+
|
3 |
+
#ifndef CMINIMAX_H
|
4 |
+
#define CMINIMAX_H
|
5 |
+
|
6 |
+
#include <iostream>
|
7 |
+
#include <vector>
|
8 |
+
|
9 |
+
const float FLOAT_MAX = 1000000.0;
|
10 |
+
const float FLOAT_MIN = -FLOAT_MAX;
|
11 |
+
|
12 |
+
namespace tools {
|
13 |
+
|
14 |
+
class CMinMaxStats {
|
15 |
+
public:
|
16 |
+
float maximum, minimum, value_delta_max;
|
17 |
+
|
18 |
+
CMinMaxStats();
|
19 |
+
~CMinMaxStats();
|
20 |
+
|
21 |
+
void set_delta(float value_delta_max);
|
22 |
+
void update(float value);
|
23 |
+
void clear();
|
24 |
+
float normalize(float value);
|
25 |
+
};
|
26 |
+
|
27 |
+
class CMinMaxStatsList {
|
28 |
+
public:
|
29 |
+
int num;
|
30 |
+
std::vector<CMinMaxStats> stats_lst;
|
31 |
+
|
32 |
+
CMinMaxStatsList();
|
33 |
+
CMinMaxStatsList(int num);
|
34 |
+
~CMinMaxStatsList();
|
35 |
+
|
36 |
+
void set_delta(float value_delta_max);
|
37 |
+
};
|
38 |
+
}
|
39 |
+
|
40 |
+
#endif
|
LightZero/lzero/mcts/ctree/common_lib/utils.cpp
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
// C++11
|
2 |
+
|
3 |
+
#include <iostream>
|
4 |
+
#include <algorithm>
|
5 |
+
|
6 |
+
#ifdef _WIN32
|
7 |
+
#include <Windows.h>
|
8 |
+
#else
|
9 |
+
#include <sys/time.h>
|
10 |
+
#endif
|
11 |
+
|
12 |
+
void get_time_and_set_rand_seed()
|
13 |
+
{
|
14 |
+
#ifdef _WIN32
|
15 |
+
FILETIME ft;
|
16 |
+
GetSystemTimeAsFileTime(&ft);
|
17 |
+
ULARGE_INTEGER uli;
|
18 |
+
uli.LowPart = ft.dwLowDateTime;
|
19 |
+
uli.HighPart = ft.dwHighDateTime;
|
20 |
+
uint64_t timestamp = (uli.QuadPart - 116444736000000000ULL) / 10000000ULL;
|
21 |
+
srand(timestamp % RAND_MAX);
|
22 |
+
#else
|
23 |
+
timeval tv;
|
24 |
+
gettimeofday(&tv, nullptr);
|
25 |
+
srand(tv.tv_usec);
|
26 |
+
#endif
|
27 |
+
}
|
LightZero/lzero/mcts/ctree/ctree_alphazero/mcts_alphazero.cpp
ADDED
@@ -0,0 +1,348 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
// This code is a Python extension implemented in C++ using the pybind11 library.
|
2 |
+
// It's a Monte Carlo Tree Search (MCTS) algorithm with modifications based on Google's AlphaZero paper.
|
3 |
+
// MCTS is an algorithm for making optimal decisions in a certain class of combinatorial problems.
|
4 |
+
// It's most famously used in board games like chess, Go, and shogi.
|
5 |
+
|
6 |
+
// The following lines include the necessary headers to facilitate the implementation of the MCTS algorithm.
|
7 |
+
#include "node_alphazero.h"
|
8 |
+
#include <cmath>
|
9 |
+
#include <map>
|
10 |
+
#include <random>
|
11 |
+
#include <vector>
|
12 |
+
#include <pybind11/pybind11.h>
|
13 |
+
#include <pybind11/stl.h>
|
14 |
+
#include <functional>
|
15 |
+
#include <iostream>
|
16 |
+
#include <memory>
|
17 |
+
#include <numeric>
|
18 |
+
|
19 |
+
// This line creates an alias for the pybind11 namespace, making it easier to reference in the code.
|
20 |
+
namespace py = pybind11;
|
21 |
+
|
22 |
+
// This part defines the MCTS class and its member variables.
|
23 |
+
// The MCTS class implements the MCTS algorithm, and its member variables store configuration values used in the algorithm.
|
24 |
+
class MCTS {
|
25 |
+
int max_moves;
|
26 |
+
int num_simulations;
|
27 |
+
double pb_c_base;
|
28 |
+
double pb_c_init;
|
29 |
+
double root_dirichlet_alpha;
|
30 |
+
double root_noise_weight;
|
31 |
+
py::object simulate_env;
|
32 |
+
|
33 |
+
// This part defines the constructor of the MCTS class.
|
34 |
+
// The constructor initializes the member variables with the provided arguments or with their default values.
|
35 |
+
public:
|
36 |
+
MCTS(int max_moves=512, int num_simulations=800,
|
37 |
+
double pb_c_base=19652, double pb_c_init=1.25,
|
38 |
+
double root_dirichlet_alpha=0.3, double root_noise_weight=0.25, py::object simulate_env=py::none())
|
39 |
+
: max_moves(max_moves), num_simulations(num_simulations),
|
40 |
+
pb_c_base(pb_c_base), pb_c_init(pb_c_init),
|
41 |
+
root_dirichlet_alpha(root_dirichlet_alpha),
|
42 |
+
root_noise_weight(root_noise_weight),
|
43 |
+
simulate_env(simulate_env) {}
|
44 |
+
|
45 |
+
// This function calculates the Upper Confidence Bound (UCB) score for a given node in the MCTS tree based on the parent node's visit count,
|
46 |
+
// the child node's visit count, and the child node's prior probability.
|
47 |
+
double _ucb_score(Node* parent, Node* child) {
|
48 |
+
double pb_c = std::log((parent->visit_count + pb_c_base + 1) / pb_c_base) + pb_c_init;
|
49 |
+
pb_c *= std::sqrt(parent->visit_count) / (child->visit_count + 1);
|
50 |
+
|
51 |
+
double prior_score = pb_c * child->prior_p;
|
52 |
+
double value_score = child->get_value();
|
53 |
+
return prior_score + value_score;
|
54 |
+
}
|
55 |
+
|
56 |
+
// This function adds Dirichlet noise to the prior probabilities of the actions of a given node to encourage exploration.
|
57 |
+
void _add_exploration_noise(Node* node) {
|
58 |
+
std::vector<int> actions;
|
59 |
+
for (const auto& kv : node->children) {
|
60 |
+
actions.push_back(kv.first);
|
61 |
+
}
|
62 |
+
|
63 |
+
std::default_random_engine generator;
|
64 |
+
std::gamma_distribution<double> distribution(root_dirichlet_alpha, 1.0);
|
65 |
+
|
66 |
+
std::vector<double> noise;
|
67 |
+
double sum = 0;
|
68 |
+
for (size_t i = 0; i < actions.size(); ++i) {
|
69 |
+
double sample = distribution(generator);
|
70 |
+
noise.push_back(sample);
|
71 |
+
sum += sample;
|
72 |
+
}
|
73 |
+
|
74 |
+
// Normalize the samples to simulate a Dirichlet distribution
|
75 |
+
for (size_t i = 0; i < noise.size(); ++i) {
|
76 |
+
noise[i] /= sum;
|
77 |
+
}
|
78 |
+
|
79 |
+
double frac = root_noise_weight;
|
80 |
+
for (size_t i = 0; i < actions.size(); ++i) {
|
81 |
+
node->children[actions[i]]->prior_p = node->children[actions[i]]->prior_p * (1 - frac) + noise[i] * frac;
|
82 |
+
}
|
83 |
+
}
|
84 |
+
// This function selects the child of a given node that has the highest UCB score among the legal actions.
|
85 |
+
std::pair<int, Node*> _select_child(Node* node, py::object simulate_env) {
|
86 |
+
int action = -1;
|
87 |
+
Node* child = nullptr;
|
88 |
+
double best_score = -9999999;
|
89 |
+
for (const auto& kv : node->children) {
|
90 |
+
int action_tmp = kv.first;
|
91 |
+
Node* child_tmp = kv.second;
|
92 |
+
|
93 |
+
py::list legal_actions_py = simulate_env.attr("legal_actions").cast<py::list>();
|
94 |
+
|
95 |
+
std::vector<int> legal_actions;
|
96 |
+
for (py::handle h : legal_actions_py) {
|
97 |
+
legal_actions.push_back(h.cast<int>());
|
98 |
+
}
|
99 |
+
|
100 |
+
if (std::find(legal_actions.begin(), legal_actions.end(), action_tmp) != legal_actions.end()) {
|
101 |
+
double score = _ucb_score(node, child_tmp);
|
102 |
+
if (score > best_score) {
|
103 |
+
best_score = score;
|
104 |
+
action = action_tmp;
|
105 |
+
child = child_tmp;
|
106 |
+
}
|
107 |
+
}
|
108 |
+
|
109 |
+
}
|
110 |
+
if (child == nullptr) {
|
111 |
+
child = node;
|
112 |
+
}
|
113 |
+
return std::make_pair(action, child);
|
114 |
+
}
|
115 |
+
|
116 |
+
// This function expands a leaf node by generating its children based on the legal actions and their prior probabilities.
|
117 |
+
double _expand_leaf_node(Node* node, py::object simulate_env, py::object policy_value_func) {
|
118 |
+
|
119 |
+
std::map<int, double> action_probs_dict;
|
120 |
+
double leaf_value;
|
121 |
+
py::tuple result = policy_value_func(simulate_env);
|
122 |
+
|
123 |
+
action_probs_dict = result[0].cast<std::map<int, double>>();
|
124 |
+
leaf_value = result[1].cast<double>();
|
125 |
+
|
126 |
+
|
127 |
+
py::list legal_actions_list = simulate_env.attr("legal_actions").cast<py::list>();
|
128 |
+
std::vector<int> legal_actions = legal_actions_list.cast<std::vector<int>>();
|
129 |
+
|
130 |
+
|
131 |
+
for (const auto& kv : action_probs_dict) {
|
132 |
+
int action = kv.first;
|
133 |
+
double prior_p = kv.second;
|
134 |
+
if (std::find(legal_actions.begin(), legal_actions.end(), action) != legal_actions.end()) {
|
135 |
+
node->children[action] = new Node(node, prior_p);
|
136 |
+
}
|
137 |
+
}
|
138 |
+
|
139 |
+
return leaf_value;
|
140 |
+
}
|
141 |
+
|
142 |
+
// This function returns the next action to take and the probabilities of each action based on the current state and the policy-value function.
|
143 |
+
std::pair<int, std::vector<double>> get_next_action(py::object state_config_for_env_reset, py::object policy_value_func, double temperature, bool sample) {
|
144 |
+
Node* root = new Node();
|
145 |
+
|
146 |
+
py::object init_state = state_config_for_env_reset["init_state"];
|
147 |
+
if (!init_state.is_none()) {
|
148 |
+
init_state = py::bytes(init_state.attr("tobytes")());
|
149 |
+
}
|
150 |
+
py::object katago_game_state = state_config_for_env_reset["katago_game_state"];
|
151 |
+
if (!katago_game_state.is_none()) {
|
152 |
+
// TODO(pu): polish efficiency
|
153 |
+
katago_game_state = py::module::import("pickle").attr("dumps")(katago_game_state);
|
154 |
+
}
|
155 |
+
simulate_env.attr("reset")(
|
156 |
+
state_config_for_env_reset["start_player_index"].cast<int>(),
|
157 |
+
init_state,
|
158 |
+
state_config_for_env_reset["katago_policy_init"].cast<bool>(),
|
159 |
+
katago_game_state
|
160 |
+
);
|
161 |
+
|
162 |
+
_expand_leaf_node(root, simulate_env, policy_value_func);
|
163 |
+
if (sample) {
|
164 |
+
_add_exploration_noise(root);
|
165 |
+
}
|
166 |
+
for (int n = 0; n < num_simulations; ++n) {
|
167 |
+
simulate_env.attr("reset")(
|
168 |
+
state_config_for_env_reset["start_player_index"].cast<int>(),
|
169 |
+
init_state,
|
170 |
+
state_config_for_env_reset["katago_policy_init"].cast<bool>(),
|
171 |
+
katago_game_state
|
172 |
+
);
|
173 |
+
simulate_env.attr("battle_mode") = simulate_env.attr("battle_mode_in_simulation_env");
|
174 |
+
_simulate(root, simulate_env, policy_value_func);
|
175 |
+
}
|
176 |
+
|
177 |
+
std::vector<std::pair<int, int>> action_visits;
|
178 |
+
for (int action = 0; action < simulate_env.attr("action_space").attr("n").cast<int>(); ++action) {
|
179 |
+
if (root->children.count(action)) {
|
180 |
+
action_visits.push_back(std::make_pair(action, root->children[action]->visit_count));
|
181 |
+
} else {
|
182 |
+
action_visits.push_back(std::make_pair(action, 0));
|
183 |
+
}
|
184 |
+
}
|
185 |
+
|
186 |
+
// Convert 'action_visits' into two separate arrays.
|
187 |
+
std::vector<int> actions;
|
188 |
+
std::vector<int> visits;
|
189 |
+
for (const auto& av : action_visits) {
|
190 |
+
actions.push_back(av.first);
|
191 |
+
visits.push_back(av.second);
|
192 |
+
}
|
193 |
+
|
194 |
+
|
195 |
+
std::vector<double> visits_d(visits.begin(), visits.end());
|
196 |
+
std::vector<double> action_probs = visit_count_to_action_distribution(visits_d, temperature);
|
197 |
+
|
198 |
+
int action;
|
199 |
+
if (sample) {
|
200 |
+
action = random_choice(actions, action_probs);
|
201 |
+
} else {
|
202 |
+
action = actions[std::distance(action_probs.begin(), std::max_element(action_probs.begin(), action_probs.end()))];
|
203 |
+
}
|
204 |
+
|
205 |
+
|
206 |
+
return std::make_pair(action, action_probs);
|
207 |
+
}
|
208 |
+
|
209 |
+
// This function performs a simulation from a given node until a leaf node is reached or a terminal state is reached.
|
210 |
+
void _simulate(Node* node, py::object simulate_env, py::object policy_value_func) {
|
211 |
+
while (!node->is_leaf()) {
|
212 |
+
int action;
|
213 |
+
std::tie(action, node) = _select_child(node, simulate_env);
|
214 |
+
if (action == -1) {
|
215 |
+
break;
|
216 |
+
}
|
217 |
+
simulate_env.attr("step")(action);
|
218 |
+
}
|
219 |
+
|
220 |
+
bool done;
|
221 |
+
int winner;
|
222 |
+
py::tuple result = simulate_env.attr("get_done_winner")();
|
223 |
+
done = result[0].cast<bool>();
|
224 |
+
winner = result[1].cast<int>();
|
225 |
+
|
226 |
+
double leaf_value;
|
227 |
+
if (!done) {
|
228 |
+
leaf_value = _expand_leaf_node(node, simulate_env, policy_value_func);
|
229 |
+
}
|
230 |
+
else {
|
231 |
+
if (simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>() == "self_play_mode") {
|
232 |
+
if (winner == -1) {
|
233 |
+
leaf_value = 0;
|
234 |
+
} else {
|
235 |
+
leaf_value = (simulate_env.attr("current_player").cast<int>() == winner) ? 1 : -1;
|
236 |
+
}
|
237 |
+
}
|
238 |
+
else if (simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>() == "play_with_bot_mode") {
|
239 |
+
if (winner == -1) {
|
240 |
+
leaf_value = 0;
|
241 |
+
} else if (winner == 1) {
|
242 |
+
leaf_value = 1;
|
243 |
+
} else if (winner == 2) {
|
244 |
+
leaf_value = -1;
|
245 |
+
}
|
246 |
+
}
|
247 |
+
}
|
248 |
+
if (simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>() == "play_with_bot_mode") {
|
249 |
+
node->update_recursive(leaf_value, simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>());
|
250 |
+
}
|
251 |
+
else if (simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>() == "self_play_mode") {
|
252 |
+
node->update_recursive(-leaf_value, simulate_env.attr("battle_mode_in_simulation_env").cast<std::string>());
|
253 |
+
}
|
254 |
+
}
|
255 |
+
|
256 |
+
|
257 |
+
|
258 |
+
|
259 |
+
|
260 |
+
private:
|
261 |
+
static std::vector<double> visit_count_to_action_distribution(const std::vector<double>& visits, double temperature) {
|
262 |
+
// Check if temperature is 0
|
263 |
+
if (temperature == 0) {
|
264 |
+
throw std::invalid_argument("Temperature cannot be 0");
|
265 |
+
}
|
266 |
+
|
267 |
+
// Check if all visit counts are 0
|
268 |
+
if (std::all_of(visits.begin(), visits.end(), [](double v){ return v == 0; })) {
|
269 |
+
throw std::invalid_argument("All visit counts cannot be 0");
|
270 |
+
}
|
271 |
+
|
272 |
+
std::vector<double> normalized_visits(visits.size());
|
273 |
+
|
274 |
+
// Divide visit counts by temperature
|
275 |
+
for (size_t i = 0; i < visits.size(); i++) {
|
276 |
+
normalized_visits[i] = visits[i] / temperature;
|
277 |
+
}
|
278 |
+
|
279 |
+
// Calculate the sum of all normalized visit counts
|
280 |
+
double sum = std::accumulate(normalized_visits.begin(), normalized_visits.end(), 0.0);
|
281 |
+
|
282 |
+
// Normalize the visit counts
|
283 |
+
for (double& visit : normalized_visits) {
|
284 |
+
visit /= sum;
|
285 |
+
}
|
286 |
+
|
287 |
+
return normalized_visits;
|
288 |
+
}
|
289 |
+
|
290 |
+
static std::vector<double> softmax(const std::vector<double>& values, double temperature) {
|
291 |
+
std::vector<double> exps;
|
292 |
+
double sum = 0.0;
|
293 |
+
// Compute the maximum value
|
294 |
+
double max_value = *std::max_element(values.begin(), values.end());
|
295 |
+
|
296 |
+
// Subtract the maximum value before exponentiation, for numerical stability
|
297 |
+
for (double v : values) {
|
298 |
+
double exp_v = std::exp((v - max_value) / temperature);
|
299 |
+
exps.push_back(exp_v);
|
300 |
+
sum += exp_v;
|
301 |
+
}
|
302 |
+
|
303 |
+
for (double& exp_v : exps) {
|
304 |
+
exp_v /= sum;
|
305 |
+
}
|
306 |
+
|
307 |
+
return exps;
|
308 |
+
}
|
309 |
+
|
310 |
+
static int random_choice(const std::vector<int>& actions, const std::vector<double>& probs) {
|
311 |
+
std::random_device rd;
|
312 |
+
std::mt19937 gen(rd());
|
313 |
+
std::discrete_distribution<> d(probs.begin(), probs.end());
|
314 |
+
return actions[d(gen)];
|
315 |
+
}
|
316 |
+
|
317 |
+
};
|
318 |
+
|
319 |
+
// This function uses pybind11 to expose the Node and MCTS classes to Python.
|
320 |
+
// This allows Python code to create and manipulate instances of these classes.
|
321 |
+
PYBIND11_MODULE(mcts_alphazero, m) {
|
322 |
+
py::class_<Node>(m, "Node")
|
323 |
+
.def(py::init([](Node* parent, float prior_p){
|
324 |
+
return new Node(parent ? parent : nullptr, prior_p);
|
325 |
+
}), py::arg("parent")=nullptr, py::arg("prior_p")=1.0)
|
326 |
+
.def_property_readonly("value", &Node::get_value)
|
327 |
+
.def("update", &Node::update)
|
328 |
+
.def("update_recursive", &Node::update_recursive)
|
329 |
+
.def("is_leaf", &Node::is_leaf)
|
330 |
+
.def("is_root", &Node::is_root)
|
331 |
+
.def("parent", &Node::get_parent)
|
332 |
+
.def_readwrite("prior_p", &Node::prior_p)
|
333 |
+
.def_readwrite("children", &Node::children)
|
334 |
+
.def("add_child", &Node::add_child)
|
335 |
+
.def_readwrite("visit_count", &Node::visit_count);
|
336 |
+
|
337 |
+
py::class_<MCTS>(m, "MCTS")
|
338 |
+
.def(py::init<int, int, double, double, double, double, py::object>(),
|
339 |
+
py::arg("max_moves")=512, py::arg("num_simulations")=800,
|
340 |
+
py::arg("pb_c_base")=19652, py::arg("pb_c_init")=1.25,
|
341 |
+
py::arg("root_dirichlet_alpha")=0.3, py::arg("root_noise_weight")=0.25, py::arg("simulate_env"))
|
342 |
+
.def("_ucb_score", &MCTS::_ucb_score)
|
343 |
+
.def("_add_exploration_noise", &MCTS::_add_exploration_noise)
|
344 |
+
.def("_select_child", &MCTS::_select_child)
|
345 |
+
.def("_expand_leaf_node", &MCTS::_expand_leaf_node)
|
346 |
+
.def("get_next_action", &MCTS::get_next_action)
|
347 |
+
.def("_simulate", &MCTS::_simulate);
|
348 |
+
}
|
LightZero/lzero/mcts/ctree/ctree_alphazero/node_alphazero.cpp
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#include "node_alphazero.h"
|
2 |
+
#include <pybind11/pybind11.h>
|
3 |
+
#include <pybind11/stl.h>
|
4 |
+
|
5 |
+
namespace py = pybind11;
|
6 |
+
|
7 |
+
PYBIND11_MODULE(node_alphazero, m) {
|
8 |
+
py::class_<Node>(m, "Node")
|
9 |
+
.def(py::init([](Node* parent, float prior_p){
|
10 |
+
return new Node(parent ? parent : nullptr, prior_p);
|
11 |
+
}), py::arg("parent")=nullptr, py::arg("prior_p")=1.0)
|
12 |
+
.def("value", &Node::get_value)
|
13 |
+
.def("update", &Node::update)
|
14 |
+
.def("update_recursive", &Node::update_recursive)
|
15 |
+
.def("is_leaf", &Node::is_leaf)
|
16 |
+
.def("is_root", &Node::is_root)
|
17 |
+
.def("parent", &Node::get_parent)
|
18 |
+
.def("children", &Node::get_children)
|
19 |
+
.def_readwrite("children", &Node::children)
|
20 |
+
.def("add_child", &Node::add_child)
|
21 |
+
.def("visit_count", &Node::get_visit_count);
|
22 |
+
}
|
LightZero/lzero/mcts/ctree/ctree_alphazero/node_alphazero.h
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#include <map>
|
2 |
+
#include <string>
|
3 |
+
#include <iostream>
|
4 |
+
#include <memory>
|
5 |
+
#include <mutex>
|
6 |
+
|
7 |
+
class Node {
|
8 |
+
public:
|
9 |
+
// Constructor, initializes a Node with a parent pointer and a prior probability
|
10 |
+
Node(Node* parent = nullptr, float prior_p = 1.0)
|
11 |
+
: parent(parent), prior_p(prior_p), visit_count(0), value_sum(0.0) {}
|
12 |
+
|
13 |
+
// Destructor, deletes all child nodes when a node is deleted to prevent memory leaks
|
14 |
+
~Node() {
|
15 |
+
for (auto& pair : children) {
|
16 |
+
delete pair.second;
|
17 |
+
}
|
18 |
+
}
|
19 |
+
|
20 |
+
// Returns the average value of the node
|
21 |
+
float get_value() {
|
22 |
+
return visit_count == 0 ? 0.0 : value_sum / visit_count;
|
23 |
+
}
|
24 |
+
|
25 |
+
// Updates the visit count and value sum of the node
|
26 |
+
void update(float value) {
|
27 |
+
visit_count++;
|
28 |
+
value_sum += value;
|
29 |
+
}
|
30 |
+
|
31 |
+
// Recursively updates the value and visit count of the node and its parent nodes
|
32 |
+
void update_recursive(float leaf_value, std::string battle_mode_in_simulation_env) {
|
33 |
+
// If the mode is "self_play_mode", the leaf_value is subtracted from the parent's value
|
34 |
+
if (battle_mode_in_simulation_env == "self_play_mode") {
|
35 |
+
update(leaf_value);
|
36 |
+
if (!is_root()) {
|
37 |
+
parent->update_recursive(-leaf_value, battle_mode_in_simulation_env);
|
38 |
+
}
|
39 |
+
}
|
40 |
+
// If the mode is "play_with_bot_mode", the leaf_value is added to the parent's value
|
41 |
+
else if (battle_mode_in_simulation_env == "play_with_bot_mode") {
|
42 |
+
update(leaf_value);
|
43 |
+
if (!is_root()) {
|
44 |
+
parent->update_recursive(leaf_value, battle_mode_in_simulation_env);
|
45 |
+
}
|
46 |
+
}
|
47 |
+
}
|
48 |
+
|
49 |
+
// Returns true if the node has no children
|
50 |
+
bool is_leaf() {
|
51 |
+
return children.empty();
|
52 |
+
}
|
53 |
+
|
54 |
+
// Returns true if the node has no parent
|
55 |
+
bool is_root() {
|
56 |
+
return parent == nullptr;
|
57 |
+
}
|
58 |
+
|
59 |
+
// Returns a pointer to the node's parent
|
60 |
+
Node* get_parent() {
|
61 |
+
return parent;
|
62 |
+
}
|
63 |
+
|
64 |
+
// Returns a map of the node's children
|
65 |
+
std::map<int, Node*> get_children() {
|
66 |
+
return children;
|
67 |
+
}
|
68 |
+
|
69 |
+
// Returns the node's visit count
|
70 |
+
int get_visit_count() {
|
71 |
+
return visit_count;
|
72 |
+
}
|
73 |
+
|
74 |
+
// Adds a child to the node
|
75 |
+
void add_child(int action, Node* node) {
|
76 |
+
children[action] = node;
|
77 |
+
}
|
78 |
+
|
79 |
+
public:
|
80 |
+
Node* parent; // Pointer to the parent node
|
81 |
+
float prior_p; // Prior probability of the node
|
82 |
+
int visit_count; // Count of visits to the node
|
83 |
+
float value_sum; // Sum of values of the node
|
84 |
+
std::map<int, Node*> children; // Map of child nodes
|
85 |
+
};
|