def act_commit(env, socl_net, agent_no, agent, record, T, Tfi, new_plan,
member):
assert isinstance(env, Env) and isinstance(agent, Agent)
assert isinstance(socl_net, SoclNet)
assert isinstance(new_plan, Plan) and isinstance(record, Record)
agent.policy_now = 'commit_f' # 添加当前行动记录
# 以owener的power为概率接受一个plan，如果是自身，概率为1
p_comit = 1
if new_plan.info['owner'] != agent_no:
p_comit = socl_net.power[new_plan.info['owner']][agent_no]['weight']
agent.renew_m_plan(new_plan, T + Tfi)
if (p_comit > uniform(0, 1)):
# P0-07 同上，覆盖计划前对原计划执行情况进行比较，并更新power
if not agent.a_plan is None:
dF = env.getValue(agent.state_now) \
- record.get_agent_record(agent_no, agent.a_plan.info['T_acpt'])["value"]
# if "commit" in agent.a_plan.info and agent.a_plan.info['commit']:
dp_f_a = new_plan.info['owner']
dP_r = agent.agent_arg['dP_r']['other']
# 仅对他人的power进行更新
if dp_f_a != agent_no:
dP = agent.agent_arg["dPower"](dF, dP_r)
d_pwr_updt_g = agent.agent_arg["d_pwr_updt_g"](
socl_net.power[dp_f_a][agent_no]['weight'], dP)
logging.debug("pwr change: from:%s,to:%s,dp:%.5s" %
(dp_f_a, agent_no, d_pwr_updt_g))
socl_net.power_delta(dp_f_a, agent_no, d_pwr_updt_g)
if d_pwr_updt_g < 0: # 如果结果不好，还要扣关系分
d_re_updt_g = d_pwr_updt_g
socl_net.relat_delta(dp_f_a, agent_no, d_re_updt_g)
else:
dp_f_a = agent_no
dP_r = agent.agent_arg['dP_r']['self']
# dP = agent.agent_arg["dPower"](dF, dP_r)
# d_pwr_updt_g = agent.agent_arg["d_pwr_updt_g"](socl_net.power[dp_f_a][agent_no]['weight'], dP)
# socl_net.power_delta(dp_f_a, agent_no, d_pwr_updt_g)
leadership_bill.leader_bill.add_record(record_type="m-plan",
gen=new_plan.sign,
user=Sign(
agent_no, T + Tfi,
'act_commit'))
assert new_plan.area.sign is not None
leadership_bill.leader_bill.add_record(record_type="m-info",
gen=new_plan.area.sign,
user=Sign(
agent_no, T + Tfi,
'act_commit'))
agent.a_plan = deepcopy(new_plan)
agent.a_plan.info['T_acpt'] = T + Tfi
agent.a_plan.info['commit'] = True
agent.a_plan.info['member'] = member
agent.policy_now = 'commit_t' # 添加当前行动记录
commited = True
else:
commited = False
logging.debug(agent.policy_now)
return socl_net, agent, commited
def act_dyjs(env, socl_net, agent_no, agent, record, T, Tfi):
assert isinstance(socl_net, SoclNet)
global_arg = arg.init_global_arg()
# 根据对自己影响的大小选择强化对象
out_power = []
agent.policy_now = 'dyjs_noaim' # 添加当前行动记录
for x in range(global_arg["Nagent"]):
if x != agent_no:
out_power.append(socl_net.power[x][agent_no]['weight'])
elif x == agent_no:
out_power.append(0)
if max(out_power) > socl_net.arg['power_thld']: # 对自己影响力大于一个阈值的人才会考虑定义角色
to_power = random_choice(norm_softmax(out_power))
for aim in range(global_arg['Nagent']):
if aim != to_power:
delta = env.arg['ACT']['dyjs']['delta_power'](
socl_net.power[to_power][aim]['weight'])
socl_net.power_delta(to_power, aim, delta)
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting="dyjs",
talking_type="dyjs",
speaker=Sign(agent_no, T + Tfi, 'act_dyjs'),
listener=Sign(aim, T + Tfi, 'act_dyjs'))
agent.policy_now = 'dyjs' # 添加当前行动记录
logging.debug(agent.policy_now)
return socl_net, agent
def _meeting_req_add_bill_record(host, member, meeting_name, ti):
for x in host:
for y in member:
if y not in host:
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting=meeting_name,
talking_type="meet_req",
speaker=Sign(x, ti - 1, "req"),
listener=Sign(y, ti, "req"))
def meeting_xtfg(env, agents, member, host, socl_net, record, T, Tfi): # 协调分工
assert isinstance(env, Env) and isinstance(socl_net, SoclNet)
assert isinstance(host, set) and isinstance(member, set)
assert host.issubset(member)
_meeting_req_add_bill_record(host, member, "xtfg", T + Tfi)
cent_weigh = socl_net.get_power_out_close_centrality()
# cent_weigh = socl_net.get_power_out_degree_centrality() # 改为od
plan_pool = []
# 获取所有host加权后的计划得分的比值
for x in member:
agents[x].meeting_now = 'xtfg' # 添加当前行动记录
for x in host:
if not agents[x].a_plan is None:
plan_v = env.arg['plan']['eval'](agents[x].a_plan.len_to_finish(
agents[x].a_plan.area.center), agents[x].a_plan.goal_value)
plan_pool.append({
"id": x,
"plan": agents[x].a_plan,
"weight": cent_weigh[x] * plan_v
})
w_sum = sum([pair['weight'] for pair in plan_pool])
sample_pool = [
plan_pool[i]['weight'] / w_sum for i in range(len(plan_pool))
]
for x in member:
to_commit_pair = plan_pool[random_choice(sample_pool)]
to_commit = to_commit_pair['plan']
socl_net, agents[x], commited = acts.act_commit(
env, socl_net, x, agents[x], record, T, Tfi, to_commit,
member) # 传入member，以便加入plan.info，后续执行时增强关系
# socl_net,agents[x] = acts.act_commit(env, socl_net, x, agents[x],record, T, Tfi, to_commit)
if commited:
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting="xtfg",
talking_type="commit_plan",
speaker=Sign(to_commit_pair['id'], T + Tfi, 'meeting_xtfg'),
listener=Sign(x, T + Tfi, 'meeting_xtfg'))
for u in member:
for v in member:
if u > v:
socl_net.relat[u][v]['weight'] = agents[u].agent_arg[
're_incr_g'](socl_net.relat[u][v]['weight'])
return agents, socl_net
def meeting_xxjl(env, agents, member, host, socl_net, record, T, Tfi): # 信息交流
assert isinstance(env, Env) and isinstance(socl_net, SoclNet)
assert isinstance(host, set) and isinstance(member, set)
assert host.issubset(member)
_meeting_req_add_bill_record(host, member, "xxjl", T + Tfi)
ret_info_pairs = []
logging.debug("memeber:%s,host:%s" % (member, host))
for x in member:
logging.debug("agent%s meeting_now" % x)
agents[x].meeting_now = 'xxjl' # 添加当前行动记录
for x in member:
if not x in host:
socl_net, agents[x] = acts.act_hqxx(env, socl_net, x, agents[x],
record, T, Tfi)
for x in member:
ret_info_pairs.append({
"agent_id":
x,
"info":
agents[x].get_latest_m_info(env.arg['meeting']['xxjl']['last_p_t'],
env.arg['meeting']['xxjl']['max_num'])
})
for x in member:
for ret_info_pair in ret_info_pairs:
y = ret_info_pair['agent_id']
infos = ret_info_pair['info']
if x == y or len(infos) < 1:
continue
agents[x].renew_m_info_list(infos, T + Tfi)
max_info_max = max([_info.info['max'] for _info in infos])
if max_info_max > env.getValue(agents[x].state_now):
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting='xxjl',
talking_type='get_useful_info',
speaker=Sign(y, T + Tfi, 'xxjl'),
listener=Sign(x, T + Tfi, 'xxjl'))
for u in member:
for v in member:
if u > v:
socl_net.relat[u][v]['weight'] = agents[u].agent_arg[
're_incr_g'](socl_net.relat[u][v]['weight'])
return agents, socl_net
def __init__(self):
# 环境初始化
self.global_arg = arg.init_global_arg()
env_arg = arg.init_env_arg(self.global_arg)
# 增加nk的一个读入操作
self.main_env = Env(env_arg)
for model_type in ['st', 'ed']:
if all_config['checkpoint']['env'][model_type]['enable']:
self.main_env.nkmodel_load(all_config['checkpoint']['env']['path'], model_type)
self.main_env.nkmodel_save(all_config["nkmodel_path"][model_type], model_type)
# 个体初始化
self.agents = []
csv_head_agent = ['agent_no'] + ['st_state'] + ['st_value'] + ['insight'] + ['xplr'] + ['xplt'] + ['enable']
moniter.AppendToCsv(csv_head_agent, all_config['agent_csv_path'])
for i in range(self.global_arg["Nagent"]):
# 个体随机初始位置
start_st_label = [randint(0, self.main_env.P - 1) for j in range(self.main_env.N)]
state_start = State(start_st_label)
self.agents.append(Agent(arg.init_agent_arg(self.global_arg,
self.main_env.arg),
self.main_env))
self.agents[i].state_now = deepcopy(state_start)
self.agents[i].agent_id = i
# 去除了一开始给一个全局area，改为添加一个包含起点的点area
start_area = Area(self.agents[i].state_now, [False] * self.main_env.N, 0)
start_area.info = get_area_sample_distr(env=self.main_env, area=start_area, state=self.agents[i].state_now,
T_stmp=0, sample_num=1, dfs_r=1)
start_area.sign = Sign(i, 0, 'start')
self.agents[i].renew_m_info(start_area, 0)
self.a_plan = None
logging.info("state:%s, st_value:%s,insight:%.5s ,xplr:%.5s, xplt:%.5s, enable:%.5s" % (
str(self.agents[i].state_now),
self.main_env.getValue(self.agents[i].state_now, 0),
self.agents[i].agent_arg['a']['insight'],
self.agents[i].agent_arg['a']['xplr'],
self.agents[i].agent_arg['a']['xplt'],
self.agents[i].agent_arg['a']['enable']))
# 记录agent信息
csv_info_agent = ['agent%d' % i] \
+ [self.agents[i].state_now] \
+ [self.main_env.getValue(self.agents[i].state_now, 0)] \
+ [self.agents[i].agent_arg['a']['insight']] \
+ [self.agents[i].agent_arg['a']['xplr']] \
+ [self.agents[i].agent_arg['a']['xplt']] \
+ [self.agents[i].agent_arg['a']['enable']]
moniter.AppendToCsv(csv_info_agent, all_config['agent_csv_path'])
# 社会网络初始化
soclnet_arg = arg.init_soclnet_arg(self.global_arg, env_arg)
self.socl_net = SoclNet(soclnet_arg)
self.socl_net.new_flat_init() # 修改初始化方法
# self.socl_net.flat_init()
if all_config['checkpoint']['socl_network']['enable']:
self.socl_net.power_load(all_config['checkpoint']['socl_network']['power'])
self.socl_net.relat_load(all_config['checkpoint']['socl_network']['relat'])
self.record = Record()
self.metric = metrics.register_all_metrics(metrics.Metrics())
def meeting_tljc(env, agents, member, host, socl_net, record, T, Tfi): # 讨论决策
assert isinstance(member, set) and isinstance(host, set)
assert host.issubset(member)
assert isinstance(agents[0], Agent)
_meeting_req_add_bill_record(host, member, "tljc", T + Tfi)
# host的max_area中，最好的区域用于制定计划
all_max_area = [agents[x].get_max_area() for x in host]
max_area = max(all_max_area, key=lambda a: a.info['max'])
for x in member:
agents[x].meeting_now = 'tljc' # 添加当前行动记录
new_plans = []
for x in member:
new_plans.append(acts._act_jhnd_get_plan(env, agents[x], max_area))
new_plans[-1].info['owner'] = x
new_plans[-1].info['T_gen'] = T + Tfi
new_plans[-1].sign = Sign(x, T + Tfi, 'tljc')
for x in member:
agents[x].renew_m_plan_list(new_plans, T + Tfi)
fin_plan = max(new_plans,
key=lambda plan: plan.goal_value) # 仅根据goal_value排序，不考虑距离
for x in member:
socl_net, agents[x], use_plan = acts.act_jhjc(env, socl_net, x,
agents[x], record, T,
Tfi, fin_plan)
if use_plan:
for h in host:
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting='tljc',
talking_type='get_a_plan',
speaker=Sign(h, T + Tfi, 'tljc'),
listener=Sign(x, T + Tfi, 'tljc'))
for u in member:
for v in member:
if u > v:
socl_net.relat[u][v]['weight'] = agents[u].agent_arg[
're_incr_g'](socl_net.relat[u][v]['weight'])
return agents, socl_net
def act_whlj(env, socl_net, agent_no, agent, record, T, Tfi):
assert isinstance(socl_net, SoclNet)
global_arg = arg.init_global_arg()
to_select = [x for x in range(global_arg['Nagent'])]
del to_select[agent_no]
to_whlj = sample(to_select, env.arg['ACT']['whlj']['k'])
for aim in to_whlj:
delta = env.arg['ACT']['whlj']['delta_relate'](
socl_net.relat[aim][agent_no]['weight'])
socl_net.relat_delta(aim, agent_no, delta)
leadership_bill.leader_bill.add_record(
record_type='talking',
meeting="whlj",
talking_type="whlj",
speaker=Sign(agent_no, T + Tfi, 'act_whlj'),
listener=Sign(aim, T + Tfi, 'act_whlj'))
agent.policy_now = 'whlj' # 添加当前行动记录
logging.debug(agent.policy_now)
return socl_net, agent
def act_jhnd(env, socl_net, agent_no, agent, record, T, Tfi): # 计划拟定
assert isinstance(env, Env) and isinstance(agent, Agent)
max_area = agent.get_max_area(
) # TODO WARNING 有可能get一个到达过的局部最优点，不停生成回到这个点的计划
new_plan = _act_jhnd_get_plan(env, agent, max_area)
new_plan.sign = Sign(agent_no, T + Tfi, 'act_jhnd')
new_plan.info['owner'] = agent_no
new_plan.info['T_gen'] = T + Tfi
agent.renew_m_plan(new_plan, Tfi)
socl_net, agent, use_plan = act_jhjc(env, socl_net, agent_no, agent,
record, T, Tfi, new_plan)
return socl_net, agent
def act_jhzx(env, socl_net, agent_no, agent, record, T, Tfi): # 计划执行
assert isinstance(agent.a_plan, Plan)
# 已经完成原地不动，在area上，向goal移动，在area外向area中心点移动
agent.state_now = agent.a_plan.next_step(agent.state_now)
# 将现有的点作为一个区域保存
new_area = Area(agent.state_now, [False] * env.N, 0)
new_area.info = get_area_sample_distr(env=env,
area=new_area,
state=agent.state_now,
T_stmp=T + Tfi,
sample_num=1,
dfs_r=1)
# NOTE cid 删除OB扰动，实际到达的点应该给一个客观值
# for k in new_area.info:
# new_area.info[k] = agent.agent_arg['ob'](new_area.info[k])
new_area.sign = Sign(agent_no, T + Tfi, "act_jhzx")
agent.renew_m_info(new_area, T + Tfi)
# 添加当前行动记录
agent.policy_now = 'jhzx'
# 计划执行完毕后，清空计划
if agent.a_plan.is_arrive(agent.state_now):
# NOTE cid 添加了清空计划前对计划的评价
if agent.a_plan is not None:
dF = env.getValue(agent.state_now) \
- record.get_agent_record(agent_no, agent.a_plan.info['T_acpt'])["value"]
# if "commit" in agent.a_plan.info and agent.a_plan.info['commit']: # 意味着来自tljc的plan就不更新了？
dp_f_a = agent.a_plan.info['owner']
dP_r = agent.agent_arg['dP_r']['other']
# 仅对他人的power进行更新
if dp_f_a != agent_no:
dP = agent.agent_arg["dPower"](dF, dP_r)
d_pwr_updt_g = agent.agent_arg["d_pwr_updt_g"](
socl_net.power[dp_f_a][agent_no]['weight'], dP)
logging.debug("pwr change: from:%s,to:%s,dp:%.5s" %
(dp_f_a, agent_no, d_pwr_updt_g))
socl_net.power_delta(dp_f_a, agent_no, d_pwr_updt_g)
if d_pwr_updt_g < 0: # 如果结果不好，还要扣关系分
d_re_updt_g = d_pwr_updt_g
socl_net.relat_delta(dp_f_a, agent_no, d_re_updt_g)
else:
dp_f_a = agent_no
dP_r = agent.agent_arg['dP_r']['self']
agent.policy_now = 'jhzx_fin'
agent.a_plan = None
logging.debug(agent.policy_now)
return socl_net, agent
def act_hqxx(env, socl_net, agent_no, agent, record, T, Tfi): # 获取信息
assert isinstance(env, Env) and isinstance(agent, Agent)
mask_t_id = sample(range(
env.N), env.arg["ACT"]["hqxx"]["mask_n"]) # 随机从N个位点中选择mask_n的位允许修改
mask_t = [False for i in range(env.N)]
for i in mask_t_id:
mask_t[i] = True
jump_d = env.arg['ACT']['hqxx']['dist']
state_t = agent.state_now
# 寻找一个包含当前节点的区域的中心，记为state_t
for i in range(jump_d):
state_t = state_t.walk(sample(mask_t_id, 1)[0], sample([-1, 1], 1)[0])
# 生成一个区域
new_area = Area(state_t, mask_t, env.arg['ACT']['hqxx']['dist'])
# 从区域中取样，获取信息，目前支持Max，Min,Avg,Mid,p0.15,p0.85
new_area.info = get_area_sample_distr(
env=env,
area=new_area,
T_stmp=T + Tfi,
state=agent.state_now,
sample_num=env.arg['ACT']['hqxx']['sample_n'],
dfs_r=env.arg['ACT']['hqxx']['dfs_p'])
# 加ob扰动
for k in new_area.info:
new_area.info[k] = agent.agent_arg['ob'](new_area.info[k])
# 把信息更新到状态中
new_area.sign = Sign(agent_no, T + Tfi, 'act_hqxx')
agent.renew_m_info(new_area, T + Tfi)
logging.debug(len(agent.frame_arg['PSM']['m-info']))
agent.policy_now = 'hqxx' # 添加当前行动记录
logging.debug(agent.policy_now)
return socl_net, agent
def act_zyzx(env, socl_net, agent_no, agent, record, T, Tfi):
global_area = Area(agent.state_now, [True] * env.N,
env.P // 2 * env.N) # 自由执行自带从global_area找下一个点
state_next = global_area.rand_walk(agent.state_now)
value_now = env.getValue(agent.state_now, T)
value_next = env.arg['ACT']['zyzx']['ob'](env.getValue(
state_next, T)) # 改为所有人是一样的ob影响参数
# agent.agent_arg['ob'](env.getValue(state_next, T))
dE = value_next - value_now
kT0 = env.arg['ACT']['xdzx']['kT0']
cd = env.arg['ACT']['xdzx']['cool_down']
# cd = env.arg['ACT']['xdzx']['cool_down'] ** (env.arg['P'] ** (env.arg['N'] - inter_area_mskn))
# 随着时间推移，容忍度越来越低 按stage来衰减
cd_T = T
tol = kT0 * cd**cd_T # 容忍度
# logging.debug("tol:{}".format(tol))
if (dE >= 0 or (tol >= 1e-10 and exp(dE / tol) > uniform(0, 1))):
agent.state_now = state_next
new_area = Area(agent.state_now, [False] * env.N, 0)
new_area.info = get_area_sample_distr(env=env,
area=new_area,
state=agent.state_now,
T_stmp=T + Tfi,
sample_num=1,
dfs_r=1)
new_area.sign = Sign(agent_no, T + Tfi, "act_zyzx")
# NOTE cid 删除OB扰动，实际到达的点应该给一个客观值
# for k in new_area.info:
# new_area.info[k] = agent.agent_arg['ob'](new_area.info[k])
agent.renew_m_info(new_area, T + Tfi)
agent.policy_now = 'zyzx' # 添加当前行动记录
logging.debug(agent.policy_now)
return socl_net, agent
def act_jhjc(env, socl_net, agent_no, agent, record, T, Tfi, new_plan):
assert isinstance(env, Env) and isinstance(agent, Agent)
assert isinstance(record, Record) and isinstance(socl_net, SoclNet)
new_plan_value = env.arg['ACT']['jhjc']["plan_eval"](
new_plan.len_to_finish(agent.state_now), new_plan.goal_value)
org_plan_value = env.getValue(
agent.state_now) # 如果没有a_plan，至少以当前位置作为plan value
if agent.a_plan is not None:
org_plan_value = env.arg['ACT']['jhjc']["plan_eval"](
agent.a_plan.len_to_finish(agent.state_now),
agent.a_plan.goal_value)
agent.policy_now = 'jhjc_old' # 添加当前行动记录，维持老计划
# logging.debug("new_v: %.5s, org_v:%.5s" % (new_plan_value, org_plan_value))
if new_plan_value >= org_plan_value:
# P0-07 在覆盖新计划前对旧计划的执行情况进行判断并据此更新SoclNet.power
if agent.a_plan is not None:
dF = env.getValue(agent.state_now) \
- record.get_agent_record(agent_no, agent.a_plan.info['T_acpt'])["value"]
# if "commit" in agent.a_plan.info and agent.a_plan.info['commit']:
dp_f_a = new_plan.info['owner']
dP_r = agent.agent_arg['dP_r']['other']
# 仅对他人的power进行更新
if dp_f_a != agent_no:
dP = agent.agent_arg["dPower"](dF, dP_r)
d_pwr_updt_g = agent.agent_arg["d_pwr_updt_g"](
socl_net.power[dp_f_a][agent_no]['weight'], dP)
logging.debug("pwr change: from:%s,to:%s,dp:%.5s" %
(dp_f_a, agent_no, d_pwr_updt_g))
socl_net.power_delta(dp_f_a, agent_no, d_pwr_updt_g)
if d_pwr_updt_g < 0: # 如果结果不好，还要扣关系分
d_re_updt_g = d_pwr_updt_g
socl_net.relat_delta(dp_f_a, agent_no, d_re_updt_g)
else:
dp_f_a = agent_no
dP_r = agent.agent_arg['dP_r']['self']
# dP = agent.agent_arg["dPower"](dF, dP_r)
# d_pwr_updt_g = agent.agent_arg["d_pwr_updt_g"](socl_net.power[dp_f_a][agent_no]['weight'], dP)
# socl_net.power_delta(dp_f_a, agent_no, d_pwr_updt_g)
new_plan.info['T_acpt'] = T + Tfi
agent.a_plan = new_plan
assert new_plan.area.sign is not None
leadership_bill.leader_bill.add_record(record_type="m-plan",
gen=new_plan.sign,
user=Sign(
agent_no, T + Tfi,
'act_jhjc'))
leadership_bill.leader_bill.add_record(record_type="m-info",
gen=new_plan.area.sign,
user=Sign(
agent_no, T + Tfi,
'act_jhjc'))
logging.debug("plan_dist: %s" %
agent.a_plan.len_to_finish(agent.state_now))
agent.policy_now = 'jhjc_new' # 添加当前行动记录,选择新计划
use_plan = True
else:
use_plan = False
logging.debug(agent.policy_now)
return socl_net, agent, use_plan