// Sample one possible sequence of future events, up to 'dfr' cycles.
reward_t SearchNode::sample(Agent &agent, unsigned int dfr) {
double newReward;
if (dfr == 0) {
return 0;
} else if (m_chance_node) {
// Generate whole observation-reward percept,
// according to the agent's model of the environment.
percept_t obs;
percept_t rew;
agent.genPerceptAndUpdate(obs, rew);
// Calculate the index of whole percept
percept_t percept = (rew << agent.numObsBits()) | obs;
if (m_child.count(percept) == 0) {
m_child[percept] = new SearchNode(false, agent.numActions());
}
newReward = rew + m_child[percept]->sample(agent, dfr - 1);
} else if (m_visits == 0) {
newReward = playout(agent, dfr);
} else {
// Select an action to sample.
action_t action = selectAction(agent, dfr);
agent.modelUpdate(action);
newReward = m_child[action]->sample(agent, dfr);
}
// Update our estimate of the future reward.
m_mean = (1.0 / (double) (m_visits + 1)) * (newReward + m_visits * m_mean);
++m_visits;
return newReward;
}
// simulate a sequence of random actions, returning the accumulated reward.
static reward_t playout(Agent &agent, unsigned int playout_len) {
reward_t r = 0;
for (unsigned int i = 0; i < playout_len; ++i) {
// Pick a random action
action_t a = agent.genRandomAction();
agent.modelUpdate(a);
// Generate a random percept distributed according to the agent's
// internal model of the environment.
percept_t rew;
percept_t obs;
agent.genPerceptAndUpdate(obs, rew);
r = r + rew;
}
return r;
}
