#define LOG(kind,...) GNUNET_log_from (kind, "ats-ril",__VA_ARGS__)
-#define RIL_MIN_BW ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__)
-#define RIL_MAX_BW GNUNET_ATS_MaxBandwidth
+#define RIL_MIN_BW (1 * ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__))
+#define RIL_MAX_BW GNUNET_ATS_MaxBandwidth
-#define RIL_ACTION_INVALID -1
-#define RIL_INTERVAL_EXPONENT 10
-#define RIL_UTILITY_MAX (double) RIL_MAX_BW
+#define RIL_ACTION_INVALID -1
+#define RIL_INTERVAL_EXPONENT 10
+#define RIL_UTILITY_DELAY_MAX 1000
-#define RIL_DEFAULT_STEP_TIME_MIN GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 500)
-#define RIL_DEFAULT_STEP_TIME_MAX GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 3000)
+#define RIL_DEFAULT_STEP_TIME_MIN GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 200)
+#define RIL_DEFAULT_STEP_TIME_MAX GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 1000)
#define RIL_DEFAULT_ALGORITHM RIL_ALGO_Q
#define RIL_DEFAULT_SELECT RIL_SELECT_EGREEDY
#define RIL_DEFAULT_WELFARE RIL_WELFARE_EGALITARIAN
#define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.1
#define RIL_DEFAULT_TRACE_DECAY 0.5
#define RIL_DEFAULT_EXPLORE_RATIO 0.1
-#define RIL_DEFAULT_RBF_DIVISOR 10
+#define RIL_DEFAULT_RBF_DIVISOR 50
#define RIL_DEFAULT_GLOBAL_REWARD_SHARE 0.5
#define RIL_DEFAULT_TEMPERATURE 1.0
#define RIL_INC_DEC_STEP_SIZE 1
-#define RIL_NOP_BONUS 0.5
+#define RIL_NOP_DECAY 0.5
/**
* ATS reinforcement learning solver
enum RIL_Action_Type
{
RIL_ACTION_NOTHING = 0,
- RIL_ACTION_BW_IN_DBL = -2, //TODO! put actions back
+ RIL_ACTION_BW_IN_DBL = -2, //TODO? Potentially add more actions
RIL_ACTION_BW_IN_HLV = -3,
RIL_ACTION_BW_IN_INC = 1,
RIL_ACTION_BW_IN_DEC = 2,
RIL_ACTION_BW_OUT_DBL = -4,
RIL_ACTION_BW_OUT_HLV = -5,
- RIL_ACTION_BW_OUT_INC = -6,
- RIL_ACTION_BW_OUT_DEC = -7,
- RIL_ACTION_TYPE_NUM = 3
+ RIL_ACTION_BW_OUT_INC = 3,
+ RIL_ACTION_BW_OUT_DEC = 4,
+ RIL_ACTION_TYPE_NUM = 5
};
enum RIL_Algorithm
*/
double ** E;
+ /**
+ * Whether to reset the eligibility traces to 0 after a Q-exploration step
+ */
+ int eligibility_reset;
+
/**
* Address in use
*/
*/
unsigned long long bw_in_assigned;
+ /**
+ * Bandwidth inbound actually utilized in the network
+ */
+ unsigned long long bw_in_utilized;
+
/**
* Total available outbound bandwidth
*/
unsigned long long bw_out_available;
/**
- * * Bandwidth outbound assigned in network after last step
+ * Bandwidth outbound assigned in network after last step
*/
unsigned long long bw_out_assigned;
+ /**
+ * Bandwidth outbound actually utilized in the network
+ */
+ unsigned long long bw_out_utilized;
+
/**
* Number of active agents in scope
*/
};
/*
- * Private functions
+ * "Private" functions
* ---------------------------
*/
* @return estimation value
*/
static double
-agent_estimate_q (struct RIL_Peer_Agent *agent, double *state, int action)
+agent_q (struct RIL_Peer_Agent *agent, double *state, int action)
{
int i;
double result = 0;
* @return wrapped address
*/
static struct RIL_Address_Wrapped *
-agent_address_get (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
+agent_address_get_wrapped (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
{
struct RIL_Address_Wrapped *cur;
}
+static int
+agent_action_is_possible (struct RIL_Peer_Agent *agent, int action)
+{
+ int address_index;
+
+ switch (action)
+ {
+ case RIL_ACTION_NOTHING:
+ return GNUNET_YES;
+ break;
+ case RIL_ACTION_BW_IN_INC:
+ case RIL_ACTION_BW_IN_DBL:
+ if (agent->bw_in >= RIL_MAX_BW)
+ return GNUNET_NO;
+ else
+ return GNUNET_YES;
+ break;
+ case RIL_ACTION_BW_IN_DEC:
+ case RIL_ACTION_BW_IN_HLV:
+ if (agent->bw_in <= RIL_MIN_BW)
+ return GNUNET_NO;
+ else
+ return GNUNET_YES;
+ break;
+ case RIL_ACTION_BW_OUT_INC:
+ case RIL_ACTION_BW_OUT_DBL:
+ if (agent->bw_out >= RIL_MAX_BW)
+ return GNUNET_NO;
+ else
+ return GNUNET_YES;
+ break;
+ case RIL_ACTION_BW_OUT_DEC:
+ case RIL_ACTION_BW_OUT_HLV:
+ if (agent->bw_out <= RIL_MIN_BW)
+ return GNUNET_NO;
+ else
+ return GNUNET_YES;
+ break;
+ default:
+ if ((action >= RIL_ACTION_TYPE_NUM) && (action < agent->n)) //switch address action
+ {
+ address_index = action - RIL_ACTION_TYPE_NUM;
+
+ GNUNET_assert(address_index >= 0);
+ GNUNET_assert(
+ address_index <= agent_address_get_index (agent, agent->addresses_tail->address_naked));
+
+ if ((agent_address_get_index(agent, agent->address_inuse) == address_index) ||
+ agent->address_inuse->active)
+ return GNUNET_NO;
+ else
+ return GNUNET_YES;
+ break;
+ }
+ // error - action does not exist
+ GNUNET_assert(GNUNET_NO);
+ }
+}
+
+
/**
* Gets the action, with the maximal estimated Q-value (i.e. the one currently estimated to bring the
* most reward in the future)
agent_get_action_max (struct RIL_Peer_Agent *agent, double *state)
{
int i;
- int num_actions;
int max_i = RIL_ACTION_INVALID;
double cur_q;
double max_q = -DBL_MAX;
- num_actions = agent->address_inuse->used ? RIL_ACTION_TYPE_NUM : agent->n;
-
- for (i = 0; i < num_actions; i++)
+ for (i = 0; i < agent->n; i++)
{
- cur_q = agent_estimate_q (agent, state, i);
- if (cur_q > max_q)
+ if (agent_action_is_possible(agent, i))
{
- max_q = cur_q;
- max_i = i;
+ cur_q = agent_q (agent, state, i);
+ if (cur_q > max_q)
+ {
+ max_q = cur_q;
+ max_i = i;
+ }
}
}
return max_i;
}
+/**
+ * Chooses a random action from the set of possible ones
+ *
+ * @param agent the agent performing the action
+ * @return the action index
+ */
+static int
+agent_get_action_random (struct RIL_Peer_Agent *agent)
+{
+ int i;
+ int is_possible[agent->n];
+ int sum = 0;
+ int r;
+
+ for (i = 0; i<agent->n; i++)
+ {
+ if (agent_action_is_possible(agent, i))
+ {
+ is_possible[i] = GNUNET_YES;
+ sum++;
+ }
+ else
+ {
+ is_possible[i] = GNUNET_NO;
+ }
+ }
+
+ r = GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK, sum);
+
+ sum = -1;
+ for (i = 0; i<agent->n; i++)
+ {
+ if (is_possible[i])
+ {
+ sum++;
+ if (sum == r)
+ return i;
+ }
+ }
+
+ GNUNET_assert(GNUNET_NO);
+}
+
/**
* Updates the weights (i.e. coefficients) of the weight vector in matrix W for action a
* @param a_prime the new
*/
static void
-agent_update_weights (struct RIL_Peer_Agent *agent, double reward, double *s_next, int a_prime)
+agent_update (struct RIL_Peer_Agent *agent, double reward, double *s_next, int a_prime)
{
int i;
+ int k;
double delta;
- double *theta = agent->W[agent->a_old];
+ double **theta = agent->W;
delta = agent->envi->global_discount_integrated * reward; //reward
- delta += agent->envi->global_discount_variable * agent_estimate_q (agent, s_next, a_prime); //discounted future value
- delta -= agent_estimate_q (agent, agent->s_old, agent->a_old); //one step
+ delta += agent->envi->global_discount_variable * agent_q (agent, s_next, a_prime); //discounted future value
+ delta -= agent_q (agent, agent->s_old, agent->a_old); //one step
// LOG(GNUNET_ERROR_TYPE_INFO, "update() Step# %llu Q(s,a): %f a: %f r: %f y: %f Q(s+1,a+1) = %f delta: %f\n",
// agent->step_count,
-// agent_estimate_q (agent, agent->s_old, agent->a_old),
+// agent_q (agent, agent->s_old, agent->a_old),
// agent->envi->parameters.alpha,
// reward,
// agent->envi->global_discount_variable,
-// agent_estimate_q (agent, s_next, a_prime),
+// agent_q (agent, s_next, a_prime),
// delta);
- for (i = 0; i < agent->m; i++)
+ for (k = 0; k < agent->n; k++)
{
-// LOG(GNUNET_ERROR_TYPE_INFO, "alpha = %f delta = %f e[%d] = %f\n",
-// agent->envi->parameters.alpha,
-// delta,
-// i,
-// agent->e[i]);
- theta[i] += agent->envi->parameters.alpha * delta * agent->s_old[i] * agent->E[a_prime][i];
+ for (i = 0; i < agent->m; i++)
+ {
+ // LOG(GNUNET_ERROR_TYPE_INFO, "alpha = %f delta = %f e[%d] = %f\n",
+ // agent->envi->parameters.alpha,
+ // delta,
+ // i,
+ // agent->e[i]);
+ theta[k][i] += agent->envi->parameters.alpha * delta * agent->E[k][i];
+ }
}
}
* Changes the eligibility trace vector e in various manners:
* #RIL_E_ACCUMULATE - adds @a feature to each component as in accumulating eligibility traces
* #RIL_E_REPLACE - resets each component to @a feature as in replacing traces
- * #RIL_E_SET - multiplies e with discount factor and lambda as in the update rule
+ * #RIL_E_DISCOUNT - multiplies e with discount factor and lambda as in the update rule
* #RIL_E_ZERO - sets e to 0 as in Watkin's Q-learning algorithm when exploring and when initializing
*
* @param agent the agent handle
agent->E[action][i] += feature[i];
break;
case RIL_E_REPLACE:
- agent->E[action][i] = (agent->envi->global_discount_variable * agent->envi->parameters.lambda * agent->E[action][i]) > feature[i] ? agent->E[action][i] : feature[i];
+ agent->E[action][i] = agent->E[action][i] > feature[i] ? agent->E[action][i] : feature[i];
break;
case RIL_E_DISCOUNT:
- agent->E[action][i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
+ for (k = 0; k < agent->n; k++)
+ {
+ agent->E[k][i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
+ }
break;
case RIL_E_ZERO:
for (k = 0; k < agent->n; k++)
}
}
-
+/**
+ * Informs the environment about the status of the solver
+ *
+ * @param solver
+ * @param op
+ * @param stat
+ */
static void
ril_inform (struct GAS_RIL_Handle *solver,
- enum GAS_Solver_Operation op,
- enum GAS_Solver_Status stat)
+ enum GAS_Solver_Operation op,
+ enum GAS_Solver_Status stat)
{
if (NULL != solver->plugin_envi->info_cb)
solver->plugin_envi->info_cb (solver->plugin_envi->info_cb_cls, op, stat, GAS_INFO_NONE);
}
-static unsigned long long
-ril_network_get_assigned (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
-{
- struct RIL_Peer_Agent *cur;
- struct RIL_Scope *net;
- unsigned long long sum = 0;
-
- for (cur = solver->agents_head; NULL != cur; cur = cur->next)
- {
- if (cur->is_active && cur->address_inuse)
- {
- net = cur->address_inuse->solver_information;
- if (net->type == type)
- {
- if (direction_in)
- sum += cur->bw_in;
- else
- sum += cur->bw_out;
- }
- }
- }
-
- return sum;
-}
-
/**
* Allocates a state vector and fills it with the features present
* @param solver the solver handle
x[1] = (double) k * (double) max_bw / (double) solver->parameters.rbf_divisor;
d[0] = x[0]-y[0];
d[1] = x[1]-y[1];
- sigma = (((double) max_bw / 2) * M_SQRT2) / (double) solver->parameters.rbf_divisor;
+ sigma = (((double) max_bw / ((double) solver->parameters.rbf_divisor + 1)) * 0.5);
f = exp(-((d[0]*d[0] + d[1]*d[1]) / (2 * sigma * sigma)));
state[m++] = f;
}
return state;
}
-///*
-// * For all networks a peer has an address in, this gets the maximum bandwidth which could
-// * theoretically be available in one of the networks. This is used for bandwidth normalization.
-// *
-// * @param agent the agent handle
-// * @param direction_in whether the inbound bandwidth should be considered. Returns the maximum outbound bandwidth if GNUNET_NO
-// */
-//static unsigned long long
-//ril_get_max_bw (struct RIL_Peer_Agent *agent, int direction_in)
-//{
-// /*
-// * get the maximum bandwidth possible for a peer, e.g. among all addresses which addresses'
-// * network could provide the maximum bandwidth if all that bandwidth was used on that one peer.
-// */
-// unsigned long long max = 0;
-// struct RIL_Address_Wrapped *cur;
-// struct RIL_Scope *net;
-//
-// for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
-// {
-// net = cur->address_naked->solver_information;
-// if (direction_in)
-// {
-// if (net->bw_in_available > max)
-// {
-// max = net->bw_in_available;
-// }
-// }
-// else
-// {
-// if (net->bw_out_available > max)
-// {
-// max = net->bw_out_available;
-// }
-// }
-// }
-// return max;
-//}
-//
-///*
-// * Get the index of the quality-property in question
-// *
-// * @param type the quality property type
-// * @return the index
-// */
-//static int
-//ril_find_property_index (uint32_t type)
-//{
-// int existing_types[] = GNUNET_ATS_QualityProperties;
-// int c;
-// for (c = 0; c < GNUNET_ATS_QualityPropertiesCount; c++)
-// if (existing_types[c] == type)
-// return c;
-// return GNUNET_SYSERR;
-//}
-//
-//static int
-//ril_get_atsi (struct ATS_Address *address, uint32_t type)
-//{
-// int c1;
-// GNUNET_assert(NULL != address);
-//
-// if ((NULL == address->atsi) || (0 == address->atsi_count))
-// return 0;
-//
-// for (c1 = 0; c1 < address->atsi_count; c1++)
-// {
-// if (ntohl (address->atsi[c1].type) == type)
-// return ntohl (address->atsi[c1].value);
-// }
-// return 0;
-//}
-//
-//
-//static double
-//envi_reward_local (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
-//{
-// const double *preferences;
-// const double *properties;
-// int prop_index;
-// double pref_match = 0;
-// double bw_norm;
-// double dl_norm;
-//
-// preferences = solver->plugin_envi->get_preferences (solver->plugin_envi->get_preference_cls,
-// &agent->peer);
-// properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
-// agent->address_inuse);
-//
-// // delay in [0,1]
-// prop_index = ril_find_property_index (GNUNET_ATS_QUALITY_NET_DELAY);
-// dl_norm = 2 - properties[prop_index]; //invert property as we want to maximize for lower latencies
-//
-// // utilization in [0,1]
-// bw_norm = (((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_IN)
-// / (double) ril_get_max_bw (agent, GNUNET_YES))
-// + ((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_OUT)
-// / (double) ril_get_max_bw (agent, GNUNET_NO))) / 2;
-//
-// // preference matching in [0,4]
-// pref_match += (preferences[GNUNET_ATS_PREFERENCE_LATENCY] * dl_norm);
-// pref_match += (preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] * bw_norm);
-//
-// // local reward in [1,2]
-// return (pref_match / 4) +1;
-//}
+/**
+ * Retrieves an ATS information value of an address
+ *
+ * @param address the address in question
+ * @param type the ATS information type
+ * @return the value
+ */
+static unsigned int
+ril_get_atsi (struct ATS_Address *address, uint32_t type)
+{
+ int c1;
+ GNUNET_assert(NULL != address);
+
+ if ((NULL == address->atsi) || (0 == address->atsi_count))
+ return GNUNET_ATS_QUALITY_NET_DELAY == type ? UINT32_MAX : 1;
+ for (c1 = 0; c1 < address->atsi_count; c1++)
+ {
+ if (ntohl (address->atsi[c1].type) == type)
+ return ntohl (address->atsi[c1].value);
+ }
+ return GNUNET_ATS_QUALITY_NET_DELAY == type ? UINT32_MAX : 1;
+}
+
+/**
+ * Returns the utility value of the connection an agent manages
+ *
+ * @param agent the agent in question
+ * @return the utility value
+ */
static double
agent_get_utility (struct RIL_Peer_Agent *agent)
{
- return (double) (agent->bw_in/RIL_MIN_BW);
+ const double *preferences;
+ double delay_atsi;
+ double delay_norm;
+ double pref_match;
+
+ preferences = agent->envi->plugin_envi->get_preferences (agent->envi->plugin_envi->get_preference_cls,
+ &agent->peer);
+
+ delay_atsi = (double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_QUALITY_NET_DELAY);
+ delay_norm = RIL_UTILITY_DELAY_MAX*exp(-delay_atsi*0.00001);
+
+ pref_match = preferences[GNUNET_ATS_PREFERENCE_LATENCY] * delay_norm;
+ pref_match += preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] *
+ sqrt((double) (agent->bw_in/RIL_MIN_BW) * (double) (agent->bw_out/RIL_MIN_BW));
+
+// return (double) (agent->bw_in/RIL_MIN_BW);
// return sqrt((double) (agent->bw_in/RIL_MIN_BW) * (double) (agent->bw_out/RIL_MIN_BW));
+ return pref_match;
}
+/**
+ * Calculates the social welfare within a network scope according to what social
+ * welfare measure is set in the configuration.
+ *
+ * @param solver the solver handle
+ * @param scope the network scope in question
+ * @return the social welfare value
+ */
static double
ril_network_get_social_welfare (struct GAS_RIL_Handle *solver, struct RIL_Scope *scope)
{
switch (solver->parameters.social_welfare)
{
case RIL_WELFARE_EGALITARIAN:
- result = RIL_UTILITY_MAX;
+ result = DBL_MAX;
for (cur = solver->agents_head; NULL != cur; cur = cur->next)
{
if (cur->is_active && cur->address_inuse && (cur->address_inuse->solver_information == scope))
{
if (cur->is_active && cur->address_inuse && (cur->address_inuse->solver_information == scope))
{
- result *= agent_get_utility(cur);
+ result *= pow(agent_get_utility(cur), 1.0 / (double) scope->agent_count);
}
}
- return pow(result, 1.0 / (double) scope->agent_count);
+ return result;
}
GNUNET_assert(GNUNET_NO);
return 1;
}
+static double
+envi_get_penalty (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
+{
+ struct RIL_Scope *net;
+ unsigned long long over_max;
+ unsigned long long over_in = 0;
+ unsigned long long over_out = 0;
+
+ net = agent->address_inuse->solver_information;
+
+ if (net->bw_in_utilized > net->bw_in_available)
+ {
+ over_in = net->bw_in_utilized - net->bw_in_available;
+ if (RIL_ACTION_BW_IN_INC == agent->a_old)
+ {
+ over_in *= 2;
+ }
+ }
+ if (net->bw_out_utilized > net->bw_out_available)
+ {
+ over_out = net->bw_out_utilized - net->bw_out_available;
+ if (RIL_ACTION_BW_OUT_INC == agent->a_old)
+ {
+ over_out *= 2;
+ }
+ }
+ over_max = GNUNET_MAX (over_in , over_out) / RIL_MIN_BW;
+
+ return -1.0 * (double) over_max;
+}
+
/**
* Gets the reward for the last performed step, which is calculated in equal
* parts from the local (the peer specific) and the global (for all peers
envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
{
struct RIL_Scope *net;
- unsigned long long overutilization;
- unsigned long long over_in = 0;
- unsigned long long over_out = 0;
double objective;
double delta;
double steady;
+ double pen_share;
+ double penalty;
+ double reward;
net = agent->address_inuse->solver_information;
- if (net->bw_in_assigned > net->bw_in_available)
- over_in = net->bw_in_assigned - net->bw_in_available;
- if (net->bw_out_assigned > net->bw_out_available)
- over_out = net->bw_out_assigned - net->bw_out_available;
- overutilization = GNUNET_MAX(over_in, over_out) / RIL_MIN_BW;
-
- objective = agent_get_utility (agent) + net->social_welfare;
+ objective = (agent_get_utility (agent) + net->social_welfare) / 2;
delta = objective - agent->objective_old;
agent->objective_old = objective;
if (delta != 0)
{
- agent->nop_bonus = abs(delta) * 0;
- }
-
- LOG(GNUNET_ERROR_TYPE_DEBUG, "utility: %f, welfare: %f, objective, overutilization: %d\n", agent_get_utility (agent), net->social_welfare, objective, overutilization);
-
- steady = (RIL_ACTION_NOTHING == agent->a_old) ? agent->nop_bonus : 0;
-
- if (0 != overutilization)
- {
- return -1.0 * (double) overutilization;
+ agent->nop_bonus = delta * RIL_NOP_DECAY;
}
else
{
- return delta + steady;
+ agent->nop_bonus *= RIL_NOP_DECAY;
}
+
+ steady = (RIL_ACTION_NOTHING == agent->a_old) ? agent->nop_bonus : 0;
+
+ pen_share = 0.5;
+ penalty = envi_get_penalty(solver, agent);
+
+ reward = delta + steady;
+ return ((1 - pen_share) * reward) + (pen_share * penalty);
}
/**
struct RIL_Address_Wrapped *cur;
int i = 0;
+ cur = agent_address_get_wrapped(agent, agent->address_inuse);
+
for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
{
if (i == address_index)
}
}
+/**
+ * Selects the next action using the e-greedy strategy. I.e. with a probability
+ * of (1-e) the action with the maximum expected return will be chosen
+ * (=> exploitation) and with probability (e) a random action will be chosen.
+ * In case the Q-learning rule is set, the function also resets the eligibility
+ * traces in the exploration case (after Watkin's Q-learning).
+ *
+ * @param agent the agent selecting an action
+ * @param state the current state-feature vector
+ * @return the action index
+ */
static int
agent_select_egreedy (struct RIL_Peer_Agent *agent, double *state)
{
int action;
- int num_actions;
double r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
UINT32_MAX) / (double) UINT32_MAX;
- num_actions = agent->address_inuse->used ? RIL_ACTION_TYPE_NUM : agent->n;
-
if (r < agent->envi->parameters.explore_ratio) //explore
{
- action = GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK, num_actions);
+ action = agent_get_action_random(agent);
if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
{
- agent_modify_eligibility(agent, RIL_E_ZERO, NULL, action);
+ agent->eligibility_reset = GNUNET_YES;
}
return action;
}
else //exploit
{
action = agent_get_action_max(agent, state);
- if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
- {
- agent_modify_eligibility(agent, RIL_E_DISCOUNT, NULL, action);
- }
return action;
}
}
{
int i;
int a_max;
- int num_actions;
double eqt[agent->n];
double p[agent->n];
double sum = 0;
double r;
- num_actions = agent->address_inuse->used ? RIL_ACTION_TYPE_NUM : agent->n;
-
a_max = agent_get_action_max(agent, state);
- for (i=0; i<num_actions; i++)
+ for (i=0; i<agent->n; i++)
{
- eqt[i] = exp(agent_estimate_q(agent,state,i) / agent->envi->parameters.temperature);
- sum += eqt[i];
+ if (agent_action_is_possible(agent, i))
+ {
+ eqt[i] = exp(agent_q(agent,state,i) / agent->envi->parameters.temperature);
+ sum += eqt[i];
+ }
}
- for (i=0; i<num_actions; i++)
+ for (i=0; i<agent->n; i++)
{
- p[i] = eqt[i]/sum;
+ if (agent_action_is_possible(agent, i))
+ {
+ p[i] = eqt[i]/sum;
+ }
+ else
+ {
+ p[i] = 0;
+ }
}
r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
UINT32_MAX) / (double) UINT32_MAX;
sum = 0;
- for (i=0; i<num_actions; i++)
+ for (i=0; i<agent->n; i++)
{
if (sum + p[i] > r)
{
if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
{
- if (i == a_max)
- agent_modify_eligibility(agent, RIL_E_DISCOUNT, NULL, i);
- else
- agent_modify_eligibility(agent, RIL_E_ZERO, NULL, -1);
+ if (i != a_max)
+ agent->eligibility_reset = GNUNET_YES;
}
return i;
}
GNUNET_assert(GNUNET_NO);
}
+/**
+ * Select the next action of an agent either according to the e-greedy strategy
+ * or the softmax strategy.
+ *
+ * @param agent the agent in question
+ * @param state the current state-feature vector
+ * @return the action index
+ */
static int
agent_select_action (struct RIL_Peer_Agent *agent, double *state)
{
s_next = envi_get_state (agent->envi, agent);
reward = envi_get_reward (agent->envi, agent);
+ if (agent->eligibility_reset)
+ {
+ agent_modify_eligibility(agent, RIL_E_ZERO, NULL, -1);
+ agent->eligibility_reset = GNUNET_NO;
+ }
+ else
+ {
+ agent_modify_eligibility (agent, RIL_E_DISCOUNT, NULL, -1);
+ }
+ if (RIL_ACTION_INVALID != agent->a_old)
+ {
+ agent_modify_eligibility (agent, agent->envi->parameters.eligibility_trace_mode, agent->s_old, agent->a_old);
+ }
+
switch (agent->envi->parameters.algorithm)
{
case RIL_ALGO_SARSA:
if (RIL_ACTION_INVALID != agent->a_old)
{
//updates weights with selected action (on-policy), if not first step
- agent_update_weights (agent, reward, s_next, a_next);
+ agent_update (agent, reward, s_next, a_next);
}
- agent_modify_eligibility (agent, RIL_E_DISCOUNT, s_next, a_next);
break;
case RIL_ALGO_Q:
if (RIL_ACTION_INVALID != agent->a_old)
{
//updates weights with best action, disregarding actually selected action (off-policy), if not first step
- agent_update_weights (agent, reward, s_next, a_max);
+ agent_update (agent, reward, s_next, a_max);
}
a_next = agent_select_action (agent, s_next);
break;
GNUNET_assert(RIL_ACTION_INVALID != a_next);
- agent_modify_eligibility (agent, agent->envi->parameters.eligibility_trace_mode, s_next, a_next);
-
GNUNET_log (GNUNET_ERROR_TYPE_INFO, "step() Step# %llu R: %f IN %llu OUT %llu A: %d\n",
agent->step_count,
reward,
agent->step_count += 1;
}
+/**
+ * Prototype of the ril_step() procedure
+ *
+ * @param solver the solver handle
+ */
static void
ril_step (struct GAS_RIL_Handle *solver);
ril_step (solver);
}
+/**
+ * Determines how much of the available bandwidth is assigned. If more is
+ * assigned than available it returns 1. The function is used to determine the
+ * step size of the adaptive stepping.
+ *
+ * @param solver the solver handle
+ * @return the ratio
+ */
static double
ril_get_used_resource_ratio (struct GAS_RIL_Handle *solver)
{
net = solver->network_entries[i];
if (net.bw_in_assigned > 0) //only consider scopes where an address is actually active
{
- sum_assigned += net.bw_in_assigned;
- sum_assigned += net.bw_out_assigned;
+ sum_assigned += net.bw_in_utilized;
+ sum_assigned += net.bw_out_utilized;
sum_available += net.bw_in_available;
sum_available += net.bw_out_available;
}
ratio = 0;
}
- return ratio > 1 ? 1 : ratio; //overutilization possible, cap at 1
+ return ratio > 1 ? 1 : ratio; //overutilization is possible, cap at 1
}
/**
return NULL ;
}
+/**
+ * Determines whether more connections are allocated in a network scope, than
+ * they would theoretically fit. This is used as a heuristic to determine,
+ * whether a new connection can be allocated or not.
+ *
+ * @param solver the solver handle
+ * @param network the network scope in question
+ * @return GNUNET_YES if there are theoretically enough resources left
+ */
static int
ril_network_is_not_full (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
{
return (net->bw_in_available > RIL_MIN_BW * address_count) && (net->bw_out_available > RIL_MIN_BW * address_count);
}
+/**
+ * Unblocks an agent for which a connection request is there, that could not
+ * be satisfied. Iterates over the addresses of the agent, if one of its
+ * addresses can now be allocated in its scope the agent is unblocked,
+ * otherwise it remains unchanged.
+ *
+ * @param solver the solver handle
+ * @param agent the agent in question
+ * @param silent
+ */
static void
ril_try_unblock_agent (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int silent)
{
if (ril_network_is_not_full(solver, net->type))
{
if (NULL == agent->address_inuse)
- envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, RIL_MIN_BW, RIL_MIN_BW, silent);
+ envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, agent->bw_in, agent->bw_out, silent);
return;
}
}
agent->address_inuse = NULL;
}
+/**
+ * Determines how much the reward needs to be discounted depending on the amount
+ * of time, which has passed since the last time-step.
+ *
+ * @param solver the solver handle
+ */
static void
ril_calculate_discount (struct GAS_RIL_Handle *solver)
{
struct GNUNET_TIME_Relative time_delta;
double tau;
- // MDP case - TODO remove when debugged
+ // MDP case - TODO! remove when debugged and test SMDP case
if (solver->simulate)
{
solver->global_discount_variable = solver->parameters.gamma;
/ (double) solver->parameters.beta;
}
+/**
+ * Count the number of active agents/connections in a network scope
+ *
+ * @param solver the solver handle
+ * @param scope the network scope in question
+ * @return the number of allocated connections
+ */
static int
ril_network_count_active_agents (struct GAS_RIL_Handle *solver, struct RIL_Scope *scope)
{
return c;
}
+/**
+ * Calculates how much bandwidth is assigned in sum in a network scope, either
+ * in the inbound or in the outbound direction.
+ *
+ * @param solver the solver handle
+ * @param type the type of the network scope in question
+ * @param direction_in GNUNET_YES if the inbound direction should be summed up,
+ * otherwise the outbound direction will be summed up
+ * @return the sum of the assigned bandwidths
+ */
+static unsigned long long
+ril_network_get_assigned (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
+{
+ struct RIL_Peer_Agent *cur;
+ struct RIL_Scope *net;
+ unsigned long long sum = 0;
+
+ for (cur = solver->agents_head; NULL != cur; cur = cur->next)
+ {
+ if (cur->is_active && cur->address_inuse)
+ {
+ net = cur->address_inuse->solver_information;
+ if (net->type == type)
+ {
+ if (direction_in)
+ sum += cur->bw_in;
+ else
+ sum += cur->bw_out;
+ }
+ }
+ }
+
+ return sum;
+}
+
+/**
+ * Calculates how much bandwidth is actually utilized in sum in a network scope,
+ * either in the inbound or in the outbound direction.
+ *
+ * @param solver the solver handle
+ * @param type the type of the network scope in question
+ * @param direction_in GNUNET_YES if the inbound direction should be summed up,
+ * otherwise the outbound direction will be summed up
+ * @return the sum of the utilized bandwidths (in bytes/second)
+ */
+static unsigned long long
+ril_network_get_utilized (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
+{
+ struct RIL_Peer_Agent *cur;
+ struct RIL_Scope *net;
+ unsigned long long sum = 0;
+
+ for (cur = solver->agents_head; NULL != cur; cur = cur->next)
+ {
+ if (cur->is_active && cur->address_inuse)
+ {
+ net = cur->address_inuse->solver_information;
+ if (net->type == type)
+ {
+ if (direction_in)
+ sum += ril_get_atsi (cur->address_inuse, GNUNET_ATS_UTILIZATION_IN);
+ else
+ sum += ril_get_atsi (cur->address_inuse, GNUNET_ATS_UTILIZATION_OUT);
+ }
+ }
+ }
+
+ return sum;
+}
+
+/**
+ * Retrieves the state of the network scope, so that its attributes are up-to-
+ * date.
+ *
+ * @param solver the solver handle
+ */
static void
ril_networks_update_state (struct GAS_RIL_Handle *solver)
{
{
net = &solver->network_entries[c];
net->bw_in_assigned = ril_network_get_assigned(solver, net->type, GNUNET_YES);
+ net->bw_in_utilized = ril_network_get_utilized(solver, net->type, GNUNET_YES);
net->bw_out_assigned = ril_network_get_assigned(solver, net->type, GNUNET_NO);
+ net->bw_out_utilized = ril_network_get_utilized(solver, net->type, GNUNET_NO);
net->agent_count = ril_network_count_active_agents(solver, net);
net->social_welfare = ril_network_get_social_welfare(solver, net);
}
offset = (double) solver->parameters.step_time_min.rel_value_us;
y = factor * pow (used_ratio, RIL_INTERVAL_EXPONENT) + offset;
- GNUNET_assert(y <= (double ) solver->parameters.step_time_max.rel_value_us);
- GNUNET_assert(y >= (double ) solver->parameters.step_time_min.rel_value_us);
+ GNUNET_assert(y <= (double) solver->parameters.step_time_max.rel_value_us);
+ GNUNET_assert(y >= (double) solver->parameters.step_time_min.rel_value_us);
time_next = GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MICROSECONDS, (unsigned long long) y);
+// LOG (GNUNET_ERROR_TYPE_INFO, "ratio: %f, factor: %f, offset: %f, y: %f\n",
+// used_ratio,
+// factor,
+// offset,
+// y);
+
if (solver->simulate)
{
time_next = GNUNET_TIME_UNIT_ZERO;
{
for (k = 0; k < agent->m; k++)
{
- agent->W[i][k] = agent->envi->parameters.alpha * (1.0 - 2.0*((double) GNUNET_CRYPTO_random_u32(GNUNET_CRYPTO_QUALITY_WEAK, UINT32_MAX)/(double)UINT32_MAX));
+ agent->W[i][k] = agent->envi->parameters.alpha * (1.0 - 2.0 * ((double) GNUNET_CRYPTO_random_u32(GNUNET_CRYPTO_QUALITY_WEAK, UINT32_MAX)/(double) UINT32_MAX));
}
}
}
agent->E[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
}
agent_w_init(agent);
+ agent->eligibility_reset = GNUNET_NO;
agent->a_old = RIL_ACTION_INVALID;
agent->s_old = GNUNET_malloc (sizeof (double) * agent->m);
agent->address_inuse = NULL;
agent->objective_old = 0;
+ agent->nop_bonus = 0;
return agent;
}
}
address_index = agent_address_get_index (agent, address);
- address_wrapped = agent_address_get (agent, address);
+ address_wrapped = agent_address_get_wrapped (agent, address);
if (NULL == address_wrapped)
{
net = address->solver_information;
- GNUNET_assert(!ril_network_is_active (s, net->type));
LOG(GNUNET_ERROR_TYPE_DEBUG,
"Address not considered by agent, address was in inactive network\n");
return;
{
if (NULL != agent->addresses_head) //if peer has an address left, use it
{
- envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, RIL_MIN_BW, RIL_MIN_BW,
- GNUNET_NO);
+ envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, agent->bw_in, agent->bw_out,
+ GNUNET_YES);
}
else
{
"API_get_preferred_address() Activated agent for peer '%s', but no address available\n",
GNUNET_i2s (peer));
}
-
return agent->address_inuse;
}
projects / oweals / gnunet.git / blobdiff