#define LOG(kind,...) GNUNET_log_from (kind, "ats-ril",__VA_ARGS__)
-#define RIL_ACTION_INVALID -1
-#define RIL_FEATURES_ADDRESS_COUNT (3 + GNUNET_ATS_QualityPropertiesCount)
-#define RIL_FEATURES_NETWORK_COUNT 4
+#define MIN_BW ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__)
-#define RIL_DEFAULT_STEP_TIME GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 3000)
-#define RIL_DEFAULT_ALGORITHM RIL_ALGO_Q
-#define RIL_DEFAULT_DISCOUNT_FACTOR 0.5
-#define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.4
-#define RIL_DEFAULT_TRACE_DECAY 0.6
-#define RIL_EXPLORE_RATIO 0.1
+#define RIL_ACTION_INVALID -1
+#define RIL_FEATURES_ADDRESS_COUNT (0)// + GNUNET_ATS_QualityPropertiesCount)
+#define RIL_FEATURES_NETWORK_COUNT 2
+#define RIL_INTERVAL_EXPONENT 10
+
+#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_ALGORITHM RIL_ALGO_SARSA
+#define RIL_DEFAULT_DISCOUNT_BETA 1
+#define RIL_DEFAULT_DISCOUNT_GAMMA 0.5
+#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_GLOBAL_REWARD_SHARE 0.5
+
+#define RIL_INC_DEC_STEP_SIZE 1
/**
* ATS reinforcement learning solver
*/
enum RIL_Action_Type
{
- RIL_ACTION_NOTHING = 0,
- RIL_ACTION_BW_IN_DBL = 1,
- RIL_ACTION_BW_IN_HLV = 2,
- RIL_ACTION_BW_IN_INC = 3,
- RIL_ACTION_BW_IN_DEC = 4,
- RIL_ACTION_BW_OUT_DBL = 5,
- RIL_ACTION_BW_OUT_HLV = 6,
- RIL_ACTION_BW_OUT_INC = 7,
- RIL_ACTION_BW_OUT_DEC = 8,
- RIL_ACTION_TYPE_NUM = 9
+ RIL_ACTION_NOTHING = -1,
+ RIL_ACTION_BW_IN_DBL = -2, //TODO! put actions back
+ RIL_ACTION_BW_IN_HLV = -3,
+ RIL_ACTION_BW_IN_INC = 0,
+ RIL_ACTION_BW_IN_DEC = 1,
+ 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 = 1
};
enum RIL_Algorithm
enum RIL_Algorithm algorithm;
/**
- * Learning discount factor in the TD-update
+ * Gradient-descent step-size
*/
- float gamma;
+ double alpha;
/**
- * Gradient-descent step-size
+ * Learning discount variable in the TD-update for semi-MDPs
+ */
+ double beta;
+
+ /**
+ * Learning discount factor in the TD-update for MDPs
*/
- float alpha;
+ double gamma;
/**
* Trace-decay factor for eligibility traces
*/
- float lambda;
+ double lambda;
+
+ /**
+ * Ratio, with what probability an agent should explore in the e-greed policy
+ */
+ double explore_ratio;
+
+ /**
+ * How big the share of the global part of the reward signal is
+ */
+ double reward_global_share;
+
+ /**
+ * Minimal interval time between steps in milliseconds
+ */
+ struct GNUNET_TIME_Relative step_time_min;
+
+ /**
+ * Maximum interval time between steps in milliseconds
+ */
+ struct GNUNET_TIME_Relative step_time_max;
};
/**
/**
* Whether the agent is active or not
*/
- int active;
+ int is_active;
/**
* Number of performed time-steps
* Outbound bandwidth assigned by the agent
*/
unsigned long long bw_out;
+
+ /**
+ * Flag whether a suggestion has to be issued
+ */
+ int suggestion_issue;
+
+ /**
+ * The address which has to be issued
+ */
+ struct ATS_Address * suggestion_address;
};
struct RIL_Network
unsigned long long bw_in_available;
/**
- * Total assigned outbound bandwidth
+ * Bandwidth inbound assigned in network after last step
*/
unsigned long long bw_in_assigned;
unsigned long long bw_out_available;
/**
- * Total assigned outbound bandwidth
+ * * Bandwidth outbound assigned in network after last step
*/
unsigned long long bw_out_assigned;
};
struct GAS_RIL_Handle
{
/**
- *
+ * The solver-plugin environment of the solver-plugin API
*/
struct GNUNET_ATS_PluginEnvironment *plugin_envi;
struct GNUNET_STATISTICS_Handle *stats;
/**
- * Number of performed time-steps
+ * Number of performed steps
*/
unsigned long long step_count;
/**
- * Interval time between steps in milliseconds //TODO? Future Work: Heterogeneous stepping among agents
+ * Timestamp for the last time-step
*/
- struct GNUNET_TIME_Relative step_time;
+ struct GNUNET_TIME_Absolute step_time_last;
/**
* Task identifier of the next time-step to be executed
*/
- GNUNET_SCHEDULER_TaskIdentifier next_step;
+ GNUNET_SCHEDULER_TaskIdentifier step_next_task_id;
+
+ /**
+ * Variable discount factor, dependent on time between steps
+ */
+ double global_discount_variable;
+
+ /**
+ * Integrated variable discount factor, dependent on time between steps
+ */
+ double global_discount_integrated;
+
+ /**
+ * State vector for networks for the current step
+ */
+ double *global_state_networks;
+
+ /**
+ * Lock for bulk operations
+ */
+ int bulk_lock;
+
+ /**
+ * Number of changes during a lock
+ */
+ int bulk_changes;
/**
* Learning parameters
*/
struct RIL_Peer_Agent * agents_head;
struct RIL_Peer_Agent * agents_tail;
+
+ /**
+ * Shutdown
+ */
+ int done;
+
+ /**
+ * Simulate steps, i.e. schedule steps immediately
+ */
+ unsigned long long simulate;
};
/*
* ---------------------------
*/
+static int
+ril_count_agents(struct GAS_RIL_Handle * solver);
+
/**
* Estimate the current action-value for state s and action a
*
result += state[i] * agent->W[action][i];
}
+ GNUNET_assert(!isnan(result));
+
+ if (isinf(result))
+ {
+ return isinf(result) * UINT32_MAX; //TODO! fix
+ }
+
return result;
}
agent_decide_exploration (struct RIL_Peer_Agent *agent)
{
//TODO? Future Work: Improve exploration/exploitation trade-off by different mechanisms than e-greedy
- /*
- * An e-greedy replacement could be based on the accuracy of the prediction of the Q-value
- */
double r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
UINT32_MAX) / (double) UINT32_MAX;
-if (r < RIL_EXPLORE_RATIO)
+ if (r < agent->envi->parameters.explore_ratio)
{
return GNUNET_YES;
}
double delta;
double *theta = agent->W[agent->a_old];
- delta = reward + agent_estimate_q (agent, s_next, a_prime)
- - agent_estimate_q (agent, agent->s_old, agent->a_old);
+ 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
+
+ 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->envi->parameters.alpha,
+ reward,
+ agent->envi->global_discount_variable,
+ agent_estimate_q (agent, s_next, a_prime),
+ delta);
+
for (i = 0; i < agent->m; i++)
{
- theta[i] += agent->envi->parameters.alpha * delta * (agent->e)[i];
+// 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[i];
}
}
* Changes the eligibility trace vector e in various manners:
* RIL_E_ACCUMULATE - adds 1 to each component as in accumulating eligibility traces
* RIL_E_REPLACE - resets each component to 1 as in replacing traces
- * RIL_E_SET - multiplies e with gamma and lambda as in the update rule
+ * RIL_E_SET - 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
* @param mod the kind of modification
*/
static void
-agent_modify_eligibility (struct RIL_Peer_Agent *agent, enum RIL_E_Modification mod)
+agent_modify_eligibility (struct RIL_Peer_Agent *agent, enum RIL_E_Modification mod, double *f)
{
int i;
double *e = agent->e;
- double gamma = agent->envi->parameters.gamma;
- double lambda = agent->envi->parameters.lambda;
for (i = 0; i < agent->m; i++)
{
switch (mod)
{
case RIL_E_ACCUMULATE:
- e[i] += 1;
+ e[i] += f[i];
break;
case RIL_E_REPLACE:
- e[i] = 1;
+ e[i] = f[i];
break;
case RIL_E_SET:
- e[i] = gamma * lambda;
+ e[i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
break;
case RIL_E_ZERO:
e[i] = 0;
}
}
+static void
+ril_inform (struct GAS_RIL_Handle *solver,
+ 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);
+}
+
/**
* Changes the active assignment suggestion of the handler and invokes the bw_changed callback to
* notify ATS of its new decision
{
int notify = GNUNET_NO;
- LOG(GNUNET_ERROR_TYPE_DEBUG, "set_active_suggestion()\n");
+ LOG(GNUNET_ERROR_TYPE_DEBUG, " set_active_suggestion() for peer '%s'\n", GNUNET_i2s (&agent->peer));
//address change
if (agent->address_inuse != new_address)
}
if (NULL != new_address)
{
- LOG(GNUNET_ERROR_TYPE_DEBUG, "set address active: %s\n", agent->active ? "yes" : "no");
- new_address->active = agent->active;
+ LOG(GNUNET_ERROR_TYPE_DEBUG, " set address active: %s\n", agent->is_active ? "yes" : "no");
+ new_address->active = agent->is_active;
new_address->assigned_bw_in.value__ = htonl (agent->bw_in);
new_address->assigned_bw_out.value__ = htonl (agent->bw_out);
}
if (new_address)
{
//activity change
- if (new_address->active != agent->active)
+ if (new_address->active != agent->is_active)
{
- new_address->active = agent->active;
+ new_address->active = agent->is_active;
+ notify |= GNUNET_YES;
}
//bw change
if (agent->bw_in != new_bw_in)
{
agent->bw_in = new_bw_in;
- new_address->assigned_bw_in.value__ = htonl (new_bw_out);
+ new_address->assigned_bw_in.value__ = htonl (new_bw_in);
notify |= GNUNET_YES;
}
if (agent->bw_out != new_bw_out)
}
}
- if (notify && agent->active && (GNUNET_NO == silent))
+ if (notify && agent->is_active && (GNUNET_NO == silent))
{
if (new_address)
{
- solver->plugin_envi->bandwidth_changed_cb (solver->plugin_envi->bw_changed_cb_cls,
- new_address);
+ LOG(GNUNET_ERROR_TYPE_DEBUG, " envi_set_active_suggestion() notify\n");
+ agent->suggestion_issue = GNUNET_YES;
+ agent->suggestion_address = new_address;
}
else if (agent->address_inuse)
{
+ //disconnect case, no new address
GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_in.value__));
GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_out.value__));
agent->bw_in = 0;
agent->bw_out = 0;
- //disconnect
- solver->plugin_envi->bandwidth_changed_cb (solver->plugin_envi->bw_changed_cb_cls,
- agent->address_inuse);
+
+ agent->suggestion_issue = GNUNET_YES;
+ agent->suggestion_address = agent->address_inuse;
}
}
agent->address_inuse = new_address;
}
+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_Network *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;
+}
+
+//static void
+//envi_state_networks (struct GAS_RIL_Handle *solver)
+//{
+// int i;
+// struct RIL_Network net;
+// int overutilized_in;
+// int overutilized_out;
+//
+// for (i = 0; i < solver->networks_count; i++)
+// {
+// net = solver->network_entries[i];
+//
+// overutilized_in = net.bw_in_assigned > net.bw_in_available;
+// overutilized_out = net.bw_out_assigned > net.bw_out_available;
+//
+// solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 0] = ((double) net.bw_in_assigned / (double) net.bw_in_available)*10;
+// solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 1] = (double) overutilized_in;
+// solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 2] = ((double) net.bw_out_assigned / (double) net.bw_out_available)*10;
+// solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 3] = (double) overutilized_out;
+// }
+//}
+
/**
* Allocates a state vector and fills it with the features present
- *
* @param solver the solver handle
+ * @param agent the agent handle
* @return pointer to the state vector
*/
static double *
envi_get_state (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
{
int i;
- int k;
- struct RIL_Network *net;
+// int k;
double *state = GNUNET_malloc (sizeof (double) * agent->m);
struct RIL_Address_Wrapped *cur_address;
- const double *properties;
+// const double *preferences;
+// const double *properties;
+ struct RIL_Network *net;
- for (i = 0; i < solver->networks_count; i++)
+ //copy global networks state
+ for (i = 0; i < solver->networks_count * RIL_FEATURES_NETWORK_COUNT; i++)
{
- net = &solver->network_entries[i];
- state[i * RIL_FEATURES_NETWORK_COUNT + 0] = (double) net->bw_in_assigned;
- state[i * RIL_FEATURES_NETWORK_COUNT + 1] = (double) net->bw_in_available;
- state[i * RIL_FEATURES_NETWORK_COUNT + 2] = (double) net->bw_out_assigned;
- state[i * RIL_FEATURES_NETWORK_COUNT + 3] = (double) net->bw_out_available;
+// state[i] = solver->global_state_networks[i];
}
+ net = agent->address_inuse->solver_information;
- i = i * RIL_FEATURES_NETWORK_COUNT; //first address feature
-
- for (cur_address = agent->addresses_head; NULL != cur_address; cur_address = cur_address->next)
+ state[0] = (double) net->bw_in_assigned / 1024; //(double) net->bw_in_available;
+ if (net->bw_in_assigned > net->bw_in_available)
{
- state[i++] = cur_address->address_naked->active;
- state[i++] = cur_address->address_naked->active ? agent->bw_in : 0;
- state[i++] = cur_address->address_naked->active ? agent->bw_out : 0;
- properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
- cur_address->address_naked);
- for (k = 0; k < GNUNET_ATS_QualityPropertiesCount; k++)
- {
- state[i++] = properties[k];
- }
+ state[1] = 1;// net->bw_in_available;
}
+ else
+ {
+ state[1] = 0;
+ }
+ LOG(GNUNET_ERROR_TYPE_INFO, "get_state() state[0] = %f\n", state[0]);
+ LOG(GNUNET_ERROR_TYPE_INFO, "get_state() state[1] = %f\n", state[1]);
- return state;
-}
+ LOG(GNUNET_ERROR_TYPE_INFO, "get_state() W / %08.3f %08.3f \\ \n", agent->W[0][0], agent->W[1][0]);
+ LOG(GNUNET_ERROR_TYPE_INFO, "get_state() W \\ %08.3f %08.3f / \n", agent->W[0][1], agent->W[1][1]);
-/**
- * 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 solver the solver handle
- * @param agent the agent handle
- * @param direction_in whether the inbound bandwidth should be considered. Returns the maximum outbound bandwidth if GNUNET_NO
- */
-static long long unsigned
-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.
- */
- int max = 0;
- struct RIL_Address_Wrapped *cur;
- struct RIL_Network *net;
- for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
+ //get peer features
+// preferences = solver->plugin_envi->get_preferences (solver->plugin_envi->get_preference_cls,
+// &agent->peer);
+// for (k = 0; k < GNUNET_ATS_PreferenceCount; k++)
+// {
+// state[i++] = preferences[k];
+// }
+
+ //get address specific features
+ for (cur_address = agent->addresses_head; NULL != cur_address; cur_address = cur_address->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;
- }
- }
+// //when changing the number of address specific state features, change RIL_FEATURES_ADDRESS_COUNT macro
+// state[i++] = cur_address->address_naked->active;
+// state[i++] = cur_address->address_naked->active ? agent->bw_in : 0;
+// state[i++] = cur_address->address_naked->active ? agent->bw_out : 0;
+// properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
+// cur_address->address_naked);
+// for (k = 0; k < GNUNET_ATS_QualityPropertiesCount; k++)
+// {
+// state[i++] = properties[k];
+// }
}
- 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;
+ 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_Network *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_global (struct GAS_RIL_Handle *solver)
+//{
+// int i;
+// struct RIL_Network net;
+// unsigned int sum_in_available = 0;
+// unsigned int sum_out_available = 0;
+// unsigned int sum_in_assigned = 0;
+// unsigned int sum_out_assigned = 0;
+// double ratio_in;
+// double ratio_out;
+//
+// for (i = 0; i < solver->networks_count; i++)
+// {
+// net = solver->network_entries[i];
+// sum_in_available += net.bw_in_available;
+// sum_in_assigned += net.bw_in_assigned;
+// sum_out_available += net.bw_out_available;
+// sum_out_assigned += net.bw_out_assigned;
+// }
+//
+// ratio_in = ((double) sum_in_assigned) / ((double) sum_in_available);
+// ratio_out = ((double) sum_out_assigned) / ((double) sum_out_available);
+//
+// // global reward in [1,2]
+// return ratio_in +1;
+// return ((ratio_in + ratio_out) / 2) + 1;
+//}
+
+//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;
+//}
+
/**
- * Gets the reward for the last performed step
+ * 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
+ * identical) reward.
*
- * @param solver solver handle
+ * @param solver the solver handle
+ * @param agent the agent handle
* @return the reward
*/
static double
envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
{
- /*
- * - Match the preferences of the peer with the current assignment
- * - Validity of the solution
- */
- const double *preferences;
- const double *properties;
- double pref_match = 0;
- double bw_norm;
struct RIL_Network *net;
- int prop_index;
-
- 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);
- prop_index = ril_find_property_index (GNUNET_ATS_QUALITY_NET_DELAY);
- pref_match += preferences[GNUNET_ATS_PREFERENCE_LATENCY] * (3 - properties[prop_index]); //invert property as we want to maximize for lower latencies
- bw_norm = GNUNET_MAX(2, (((
- ((double) agent->bw_in / (double) ril_get_max_bw(agent, GNUNET_YES)) +
- ((double) agent->bw_out / (double) ril_get_max_bw(agent, GNUNET_NO))
- ) / 2
- ) + 1));
- pref_match += preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] * bw_norm;
-
+// double reward = 0;
+ long long overutilized_in = 0;
+// long long overutilized_out;
+ long long assigned_in = 0;
+// long long assigned_out = 0;
+// long long unused;
+
+ //punish overutilization
net = agent->address_inuse->solver_information;
- if ((net->bw_in_assigned > net->bw_in_available) || (net->bw_out_assigned > net->bw_out_available))
+
+ if (net->bw_in_assigned > net->bw_in_available)
{
- return -1;
+ overutilized_in = (net->bw_in_assigned - net->bw_in_available);
+ assigned_in = net->bw_in_available;
}
-
- return pref_match;
+ else
+ {
+ assigned_in = net->bw_in_assigned;
+ }
+// if (net->bw_out_assigned > net->bw_out_available)
+// {
+// overutilized_out = (net->bw_out_assigned - net->bw_out_available);
+// assigned_out = net->bw_out_available;
+// }
+// else
+// {
+// assigned_out = net->bw_out_assigned;
+// }
+
+// unused = net->bw_in_available - net->bw_in_assigned;
+// unused = unused < 0 ? unused : -unused;
+
+ return (double) (assigned_in - overutilized_in) / 1024;
+
+// reward += envi_reward_global (solver) * (solver->parameters.reward_global_share);
+// reward += envi_reward_local (solver, agent) * (1 - solver->parameters.reward_global_share);
+//
+// return (reward - 1.) * 100;
}
/**
struct RIL_Peer_Agent *agent,
int direction_in)
{
+ unsigned long long new_bw;
+
if (direction_in)
{
- envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in * 2,
+ new_bw = agent->bw_in * 2;
+ if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
+ new_bw = GNUNET_ATS_MaxBandwidth;
+ envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
agent->bw_out, GNUNET_NO);
}
else
{
+ new_bw = agent->bw_out * 2;
+ if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
+ new_bw = GNUNET_ATS_MaxBandwidth;
envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
- agent->bw_out * 2, GNUNET_NO);
+ new_bw, GNUNET_NO);
}
}
struct RIL_Peer_Agent *agent,
int direction_in)
{
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
unsigned long long new_bw;
if (direction_in)
{
new_bw = agent->bw_in / 2;
- if (new_bw < min_bw)
- new_bw = min_bw;
+ if (new_bw < MIN_BW || new_bw > agent->bw_in)
+ new_bw = MIN_BW;
envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
GNUNET_NO);
}
else
{
new_bw = agent->bw_out / 2;
- if (new_bw < min_bw)
- new_bw = min_bw;
+ if (new_bw < MIN_BW || new_bw > agent->bw_out)
+ new_bw = MIN_BW;
envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
GNUNET_NO);
}
static void
envi_action_bw_inc (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
{
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
+ unsigned long long new_bw;
if (direction_in)
{
- envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in + (5 * min_bw),
+ new_bw = agent->bw_in + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
+ if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
+ new_bw = GNUNET_ATS_MaxBandwidth;
+ envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
agent->bw_out, GNUNET_NO);
}
else
{
+ new_bw = agent->bw_out + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
+ if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
+ new_bw = GNUNET_ATS_MaxBandwidth;
envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
- agent->bw_out + (5 * min_bw), GNUNET_NO);
+ new_bw, GNUNET_NO);
}
}
static void
envi_action_bw_dec (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
{
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
unsigned long long new_bw;
if (direction_in)
{
- new_bw = agent->bw_in - (5 * min_bw);
- if (new_bw < min_bw)
- new_bw = min_bw;
+ new_bw = agent->bw_in - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
+ if (new_bw < MIN_BW || new_bw > agent->bw_in)
+ new_bw = MIN_BW;
envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
GNUNET_NO);
}
else
{
- new_bw = agent->bw_out - (5 * min_bw);
- if (new_bw < min_bw)
- new_bw = min_bw;
+ new_bw = agent->bw_out - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
+ if (new_bw < MIN_BW || new_bw > agent->bw_out)
+ new_bw = MIN_BW;
envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
GNUNET_NO);
}
/**
* Puts the action into effect by calling the according function
*
- * @param solver solver handle
- * @param action action to perform by the solver
+ * @param solver the solver handle
+ * @param agent the action handle
+ * @param action the action to perform by the solver
*/
static void
envi_do_action (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int action)
agent_step (struct RIL_Peer_Agent *agent)
{
int a_next = RIL_ACTION_INVALID;
+ int explore;
double *s_next;
double reward;
+ LOG(GNUNET_ERROR_TYPE_DEBUG, " agent_step() Peer '%s', algorithm %s\n",
+ GNUNET_i2s (&agent->peer),
+ agent->envi->parameters.algorithm ? "Q" : "SARSA");
+
s_next = envi_get_state (agent->envi, agent);
reward = envi_get_reward (agent->envi, agent);
-
- LOG(GNUNET_ERROR_TYPE_DEBUG, "agent_step() with algorithm %s\n",
- agent->envi->parameters.algorithm ? "Q" : "SARSA");
+ explore = agent_decide_exploration (agent);
switch (agent->envi->parameters.algorithm)
{
case RIL_ALGO_SARSA:
- agent_modify_eligibility (agent, RIL_E_SET);
- if (agent_decide_exploration (agent))
+ if (explore)
{
a_next = agent_get_action_explore (agent, s_next);
}
{
//updates weights with selected action (on-policy), if not first step
agent_update_weights (agent, reward, s_next, a_next);
+ agent_modify_eligibility (agent, RIL_E_SET, s_next);
}
break;
//updates weights with best action, disregarding actually selected action (off-policy), if not first step
agent_update_weights (agent, reward, s_next, a_next);
}
- if (agent_decide_exploration (agent))
+ if (explore)
{
a_next = agent_get_action_explore (agent, s_next);
- agent_modify_eligibility (agent, RIL_E_ZERO);
+ agent_modify_eligibility (agent, RIL_E_ZERO, NULL);
}
else
{
a_next = agent_get_action_best (agent, s_next);
- agent_modify_eligibility (agent, RIL_E_SET);
+ agent_modify_eligibility (agent, RIL_E_SET, s_next);
}
break;
}
GNUNET_assert(RIL_ACTION_INVALID != a_next);
- agent_modify_eligibility (agent, RIL_E_ACCUMULATE);
+ agent_modify_eligibility (agent, RIL_E_ACCUMULATE, s_next);
+
+ GNUNET_log (GNUNET_ERROR_TYPE_INFO, "step() Step# %llu R: %f IN %llu OUT %llu A: %d\n",
+ agent->step_count,
+ reward,
+ agent->bw_in/1024,
+ agent->bw_out/1024,
+ a_next);
envi_do_action (agent->envi, agent, a_next);
agent->step_count += 1;
}
+static void
+ril_step (struct GAS_RIL_Handle *solver);
+
/**
- * Cycles through all agents and lets the active ones do a step. Schedules the next step.
+ * Task for the scheduler, which performs one step and lets the solver know that
+ * no further step is scheduled.
*
- * @param solver the solver handle
- * @param tc task context for the scheduler
+ * @param cls the solver handle
+ * @param tc the task context for the scheduler
*/
static void
-ril_periodic_step (void *cls, const struct GNUNET_SCHEDULER_TaskContext *tc)
+ril_step_scheduler_task (void *cls, const struct GNUNET_SCHEDULER_TaskContext *tc)
{
struct GAS_RIL_Handle *solver = cls;
+
+ solver->step_next_task_id = GNUNET_SCHEDULER_NO_TASK;
+ ril_step (solver);
+}
+
+static double
+ril_get_used_resource_ratio (struct GAS_RIL_Handle *solver)
+{
+ int i;
+ struct RIL_Network net;
+ unsigned long long sum_assigned = 0;
+ unsigned long long sum_available = 0;
+ double ratio;
+
+ for (i = 0; i < solver->networks_count; i++)
+ {
+ 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_available += net.bw_in_available;
+ sum_available += net.bw_out_available;
+ }
+ }
+ if (sum_available > 0)
+ {
+ ratio = ((double) sum_assigned) / ((double) sum_available);
+ }
+ else
+ {
+ ratio = 0;
+ }
+
+ return ratio > 1 ? 1 : ratio; //overutilization possible, cap at 1
+}
+
+/**
+ * Lookup network struct by type
+ *
+ * @param s the solver handle
+ * @param type the network type
+ * @return the network struct
+ */
+static struct RIL_Network *
+ril_get_network (struct GAS_RIL_Handle *s, uint32_t type)
+{
+ int i;
+
+ for (i = 0; i < s->networks_count; i++)
+ {
+ if (s->network_entries[i].type == type)
+ {
+ return &s->network_entries[i];
+ }
+ }
+ return NULL ;
+}
+
+static int
+ril_network_is_not_full (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
+{
+ struct RIL_Network *net;
+ struct RIL_Peer_Agent *agent;
+ unsigned long long address_count = 0;
+
+ for (agent = solver->agents_head; NULL != agent; agent = agent->next)
+ {
+ if (agent->address_inuse && agent->is_active)
+ {
+ net = agent->address_inuse->solver_information;
+ if (net->type == network)
+ {
+ address_count++;
+ }
+ }
+ }
+
+ net = ril_get_network (solver, network);
+ return (net->bw_in_available > MIN_BW * address_count) && (net->bw_out_available > MIN_BW * address_count);
+}
+
+static void
+ril_try_unblock_agent (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int silent)
+{
+ struct RIL_Address_Wrapped *addr_wrap;
+ struct RIL_Network *net;
+
+ for (addr_wrap = agent->addresses_head; NULL != addr_wrap; addr_wrap = addr_wrap->next)
+ {
+ net = addr_wrap->address_naked->solver_information;
+ if (ril_network_is_not_full(solver, net->type))
+ {
+ if (NULL == agent->address_inuse)
+ envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, MIN_BW, MIN_BW, silent);
+ return;
+ }
+ }
+ agent->address_inuse = NULL;
+}
+
+static void
+ril_calculate_discount (struct GAS_RIL_Handle *solver)
+{
+ struct GNUNET_TIME_Absolute time_now;
+ struct GNUNET_TIME_Relative time_delta;
+ double tau;
+
+ // MDP case - remove when debugged
+ if (solver->simulate)
+ {
+ solver->global_discount_variable = solver->parameters.gamma;
+ solver->global_discount_integrated = 1;
+ return;
+ }
+
+ // semi-MDP case
+
+ //calculate tau, i.e. how many real valued time units have passed, one time unit is one minimum time step
+ time_now = GNUNET_TIME_absolute_get ();
+ time_delta = GNUNET_TIME_absolute_get_difference (solver->step_time_last, time_now);
+ solver->step_time_last = time_now;
+ tau = (double) time_delta.rel_value_us
+ / (double) solver->parameters.step_time_min.rel_value_us;
+
+ //calculate reward discounts (once per step for all agents)
+ solver->global_discount_variable = pow (M_E, ((-1.) * ((double) solver->parameters.beta) * tau));
+ solver->global_discount_integrated = (1. - solver->global_discount_variable)
+ / (double) solver->parameters.beta;
+}
+
+static void
+ril_calculate_assigned_bwnet (struct GAS_RIL_Handle *solver)
+{
+ int c;
+ struct RIL_Network *net;
+
+ for (c = 0; c < solver->networks_count; c++)
+ {
+ net = &solver->network_entries[c];
+ net->bw_in_assigned = ril_network_get_assigned(solver, net->type, GNUNET_YES);
+ net->bw_out_assigned = ril_network_get_assigned(solver, net->type, GNUNET_NO);
+ }
+}
+
+/**
+ * Schedules the next global step in an adaptive way. The more resources are
+ * left, the earlier the next step is scheduled. This serves the reactivity of
+ * the solver to changed inputs.
+ *
+ * @param solver the solver handle
+ */
+static void
+ril_step_schedule_next (struct GAS_RIL_Handle *solver)
+{
+ double used_ratio;
+ double factor;
+ double y;
+ double offset;
+ struct GNUNET_TIME_Relative time_next;
+
+ used_ratio = ril_get_used_resource_ratio (solver);
+
+ GNUNET_assert(
+ solver->parameters.step_time_min.rel_value_us
+ <= solver->parameters.step_time_max.rel_value_us);
+
+ factor = (double) GNUNET_TIME_relative_subtract (solver->parameters.step_time_max,
+ solver->parameters.step_time_min).rel_value_us;
+ 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);
+
+ time_next = GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MICROSECONDS, (unsigned long long) y);
+
+ if (solver->simulate)
+ {
+ time_next = GNUNET_TIME_UNIT_ZERO;
+ }
+
+ if ((GNUNET_SCHEDULER_NO_TASK == solver->step_next_task_id) && (GNUNET_NO == solver->done))
+ {
+ solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (time_next, &ril_step_scheduler_task,
+ solver);
+ }
+}
+
+/**
+ * Triggers one step per agent
+ * @param solver
+ */
+static void
+ril_step (struct GAS_RIL_Handle *solver)
+{
struct RIL_Peer_Agent *cur;
- LOG(GNUNET_ERROR_TYPE_DEBUG, "RIL step number %d\n", solver->step_count);
+ if (GNUNET_YES == solver->bulk_lock)
+ {
+ solver->bulk_changes++;
+ return;
+ }
+
+ ril_inform (solver, GAS_OP_SOLVE_START, GAS_STAT_SUCCESS);
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, " RIL step number %d\n", solver->step_count);
+ if (0 == solver->step_count)
+ {
+ solver->step_time_last = GNUNET_TIME_absolute_get ();
+ }
+
+ ril_calculate_discount (solver);
+ ril_calculate_assigned_bwnet (solver);
+
+ //calculate network state vector
+// envi_state_networks(solver);
+
+ //trigger one step per active, unblocked agent
for (cur = solver->agents_head; NULL != cur; cur = cur->next)
{
- if (cur->active && cur->address_inuse)
+ if (cur->is_active)
{
- agent_step (cur);
+ if (NULL == cur->address_inuse)
+ {
+ ril_try_unblock_agent(solver, cur, GNUNET_NO);
+ }
+ if (cur->address_inuse)
+ {
+ agent_step (cur);
+ }
}
}
+ ril_calculate_assigned_bwnet (solver);
+
solver->step_count += 1;
- solver->next_step = GNUNET_SCHEDULER_add_delayed (solver->step_time, &ril_periodic_step, solver);
+ ril_step_schedule_next (solver);
+
+ ril_inform (solver, GAS_OP_SOLVE_STOP, GAS_STAT_SUCCESS);
+
+ ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_START, GAS_STAT_SUCCESS);
+ for (cur = solver->agents_head; NULL != cur; cur = cur->next)
+ {
+ if (cur->suggestion_issue) {
+ solver->plugin_envi->bandwidth_changed_cb(solver->plugin_envi->bw_changed_cb_cls, cur->suggestion_address);
+ cur->suggestion_issue = GNUNET_NO;
+ }
+ }
+ ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_STOP, GAS_STAT_SUCCESS);
+}
+
+static int
+ril_count_agents (struct GAS_RIL_Handle *solver)
+{
+ int c = 0;
+ struct RIL_Peer_Agent *cur_agent;
+
+ for (cur_agent = solver->agents_head; NULL != cur_agent; cur_agent = cur_agent->next)
+ {
+ c++;
+ }
+ return c;
+}
+
+static void
+agent_w_start (struct RIL_Peer_Agent *agent)
+{
+ int count;
+ struct RIL_Peer_Agent *other;
+ int i;
+ int k;
+
+ count = ril_count_agents(agent->envi);
+
+ for (i = 0; i < agent->n; i++)
+ {
+ for (k = 0; k < agent->m; k++)
+ {
+ if (0 == count) {
+ agent->W[i][k] = 1;//.1 - ((double) GNUNET_CRYPTO_random_u32(GNUNET_CRYPTO_QUALITY_WEAK, UINT32_MAX/5)/(double)UINT32_MAX);
+ }
+ else {
+ for (other = agent->envi->agents_head; NULL != other; other = other->next)
+ {
+ agent->W[i][k] += (other->W[i][k] / (double) count);
+ }
+ }
+
+ GNUNET_assert(!isinf(agent->W[i][k]));
+ }
+ }
}
/**
agent->envi = solver;
agent->peer = *peer;
agent->step_count = 0;
- agent->active = GNUNET_NO;
+ agent->is_active = GNUNET_NO;
+ agent->bw_in = MIN_BW;
+ agent->bw_out = MIN_BW;
+ agent->suggestion_issue = GNUNET_NO;
agent->n = RIL_ACTION_TYPE_NUM;
- agent->m = solver->networks_count * RIL_FEATURES_NETWORK_COUNT;
+ agent->m = (RIL_FEATURES_NETWORK_COUNT);// + GNUNET_ATS_PreferenceCount;
agent->W = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
for (i = 0; i < agent->n; i++)
{
agent->W[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
}
+ agent_w_start(agent);
agent->a_old = RIL_ACTION_INVALID;
- agent->s_old = envi_get_state (solver, agent);
+ agent->s_old = GNUNET_malloc (sizeof (double) * agent->m);
agent->e = (double *) GNUNET_malloc (sizeof (double) * agent->m);
- agent_modify_eligibility (agent, RIL_E_ZERO);
-
- GNUNET_CONTAINER_DLL_insert_tail(solver->agents_head, solver->agents_tail, agent);
+ agent_modify_eligibility (agent, RIL_E_ZERO, NULL);
return agent;
}
/**
* Deallocate agent
*
- * @param s solver handle
+ * @param solver the solver handle
* @param agent the agent to retire
*/
static void
/**
* Returns the agent for a peer
*
- * @param s solver handle
- * @param peer identity of the peer
- * @param create whether to create an agent if none is allocated yet
- * @return agent
+ * @param solver the solver handle
+ * @param peer the identity of the peer
+ * @param create whether or not to create an agent, if none is allocated yet
+ * @return the agent
*/
static struct RIL_Peer_Agent *
ril_get_agent (struct GAS_RIL_Handle *solver, const struct GNUNET_PeerIdentity *peer, int create)
if (create)
{
- return agent_init (solver, peer);
- }
- return NULL ;
-}
-
-/**
- * Lookup network struct by type
- *
- * @param s the solver handle
- * @param type the network type
- * @return the network struct
- */
-static struct RIL_Network *
-ril_get_network (struct GAS_RIL_Handle *s, uint32_t type)
-{
- int i;
-
- for (i = 0; i < s->networks_count; i++)
- {
- if (s->network_entries[i].type == type)
- {
- return &s->network_entries[i];
- }
+ cur = agent_init (solver, peer);
+ GNUNET_CONTAINER_DLL_insert_tail(solver->agents_head, solver->agents_tail, cur);
+ return cur;
}
return NULL ;
}
ril_network_is_active (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
{
struct RIL_Network *net;
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
net = ril_get_network (solver, network);
- if (net->bw_out_available < min_bw)
- return GNUNET_NO;
- return GNUNET_YES;
+ return net->bw_out_available >= MIN_BW;
}
/**
* @param pref_rel the normalized preference value for this kind over all clients
*/
void
-GAS_ril_address_change_preference (void *s,
+GAS_ril_address_change_preference (void *solver,
const struct GNUNET_PeerIdentity *peer,
enum GNUNET_ATS_PreferenceKind kind,
double pref_rel)
LOG(GNUNET_ERROR_TYPE_DEBUG,
"API_address_change_preference() Preference '%s' for peer '%s' changed to %.2f \n",
GNUNET_ATS_print_preference_type (kind), GNUNET_i2s (peer), pref_rel);
- /*
- * Nothing to do here. Preferences are considered during reward calculation.
- */
+
+ ril_step (solver);
}
/**
struct GAS_RIL_Handle *solver = GNUNET_new (struct GAS_RIL_Handle);
struct RIL_Network * cur;
int c;
- unsigned long long tmp;
char *string;
LOG(GNUNET_ERROR_TYPE_DEBUG, "API_init() Initializing RIL solver\n");
GNUNET_assert(NULL != env->get_property);
if (GNUNET_OK
- != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME", &solver->step_time))
+ != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MIN",
+ &solver->parameters.step_time_min))
{
- solver->step_time = RIL_DEFAULT_STEP_TIME;
+ solver->parameters.step_time_min = RIL_DEFAULT_STEP_TIME_MIN;
}
- if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_ALGORITHM", &string)
- && NULL != string && 0 == strcmp (string, "SARSA"))
+ if (GNUNET_OK
+ != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MAX",
+ &solver->parameters.step_time_max))
{
- solver->parameters.algorithm = RIL_ALGO_SARSA;
+ solver->parameters.step_time_max = RIL_DEFAULT_STEP_TIME_MAX;
+ }
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_ALGORITHM", &string))
+ {
+ solver->parameters.algorithm = !strcmp (string, "SARSA") ? RIL_ALGO_SARSA : RIL_ALGO_Q;
+ GNUNET_free (string);
}
else
{
solver->parameters.algorithm = RIL_DEFAULT_ALGORITHM;
}
- if (GNUNET_OK
- == GNUNET_CONFIGURATION_get_value_size (env->cfg, "ats", "RIL_DISCOUNT_FACTOR", &tmp))
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_BETA", &string))
{
- solver->parameters.gamma = (double) tmp / 100;
+ solver->parameters.beta = strtod (string, NULL);
+ GNUNET_free (string);
}
else
{
- solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_FACTOR;
+ solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
+ }
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_GAMMA", &string))
+ {
+ solver->parameters.gamma = strtod (string, NULL);
+ GNUNET_free (string);
+ }
+ else
+ {
+ solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
}
if (GNUNET_OK
- == GNUNET_CONFIGURATION_get_value_size (env->cfg, "ats", "RIL_GRADIENT_STEP_SIZE", &tmp))
+ == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GRADIENT_STEP_SIZE", &string))
{
- solver->parameters.alpha = (double) tmp / 100;
+ solver->parameters.alpha = strtod (string, NULL);
+ GNUNET_free (string);
}
else
{
solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
}
- if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_size (env->cfg, "ats", "RIL_TRACE_DECAY", &tmp))
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_TRACE_DECAY", &string))
{
- solver->parameters.lambda = (double) tmp / 100;
+ solver->parameters.lambda = strtod (string, NULL);
+ GNUNET_free (string);
}
else
{
solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
}
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_EXPLORE_RATIO", &string))
+ {
+ solver->parameters.explore_ratio = strtod (string, NULL);
+ GNUNET_free (string);
+ }
+ else
+ {
+ solver->parameters.explore_ratio = RIL_DEFAULT_EXPLORE_RATIO;
+ }
+ if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GLOBAL_REWARD_SHARE", &string))
+ {
+ solver->parameters.reward_global_share = strtod (string, NULL);
+ GNUNET_free (string);
+ }
+ else
+ {
+ solver->parameters.reward_global_share = RIL_DEFAULT_GLOBAL_REWARD_SHARE;
+ }
+ if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number (env->cfg, "ats", "RIL_SIMULATE", &solver->simulate))
+ {
+ solver->simulate = GNUNET_NO;
+ }
env->sf.s_add = &GAS_ril_address_add;
env->sf.s_address_update_property = &GAS_ril_address_property_changed;
solver->networks_count = env->network_count;
solver->network_entries = GNUNET_malloc (env->network_count * sizeof (struct RIL_Network));
solver->step_count = 0;
+ solver->global_state_networks = GNUNET_malloc (solver->networks_count * RIL_FEATURES_NETWORK_COUNT * sizeof (double));
+ solver->done = GNUNET_NO;
for (c = 0; c < env->network_count; c++)
{
cur = &solver->network_entries[c];
cur->type = env->networks[c];
cur->bw_in_available = env->in_quota[c];
- cur->bw_in_assigned = 0;
cur->bw_out_available = env->out_quota[c];
- cur->bw_out_assigned = 0;
+ LOG(GNUNET_ERROR_TYPE_INFO, "init() Quotas for %s network: IN %llu - OUT %llu\n", GNUNET_ATS_print_network_type(cur->type), cur->bw_in_available/1024, cur->bw_out_available/1024);
}
- solver->next_step = GNUNET_SCHEDULER_add_delayed (
- GNUNET_TIME_relative_multiply (GNUNET_TIME_relative_get_millisecond_ (), 1000),
- &ril_periodic_step, solver);
+ LOG(GNUNET_ERROR_TYPE_INFO, "init() Parameters:\n");
+ LOG(GNUNET_ERROR_TYPE_INFO, "init() Algorithm = %s, alpha = %f, beta = %f, lambda = %f\n",
+ solver->parameters.algorithm ? "Q" : "SARSA",
+ solver->parameters.alpha,
+ solver->parameters.beta,
+ solver->parameters.lambda);
+ LOG(GNUNET_ERROR_TYPE_INFO, "init() explore = %f, global_share = %f\n",
+ solver->parameters.explore_ratio,
+ solver->parameters.reward_global_share);
return solver;
}
LOG(GNUNET_ERROR_TYPE_DEBUG, "API_done() Shutting down RIL solver\n");
+ s->done = GNUNET_YES;
+
cur_agent = s->agents_head;
while (NULL != cur_agent)
{
cur_agent = next_agent;
}
- GNUNET_SCHEDULER_cancel (s->next_step);
+ if (GNUNET_SCHEDULER_NO_TASK != s->step_next_task_id)
+ {
+ GNUNET_SCHEDULER_cancel (s->step_next_task_id);
+ }
GNUNET_free(s->network_entries);
+ GNUNET_free(s->global_state_networks);
GNUNET_free(s);
- return NULL ;
+ return NULL;
}
/**
unsigned int n_old;
int i;
unsigned int zero;
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
+
+ LOG (GNUNET_ERROR_TYPE_DEBUG, "API_address_add()\n");
net = ril_get_network (s, network);
address->solver_information = net;
if (!ril_network_is_active (s, network))
{
LOG(GNUNET_ERROR_TYPE_DEBUG,
- "API_address_add() Did not add %s address %p for peer '%s', network does not have enough bandwidth\n",
+ "API_address_add() Did not add %s address %s for peer '%s', network does not have enough bandwidth\n",
address->plugin, address->addr, GNUNET_i2s (&address->peer));
return;
}
agent->m = m_old;
GNUNET_array_grow(agent->e, agent->m, m_new);
- if (NULL == agent->address_inuse)
- {
- net->bw_in_assigned += min_bw;
- net->bw_out_assigned += min_bw;
- envi_set_active_suggestion (s, agent, address, min_bw, min_bw, GNUNET_NO);
- }
+ ril_try_unblock_agent(s, agent, GNUNET_NO);
+
+ ril_step (s);
- LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %p for peer '%s'\n",
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %s for peer '%s'\n",
address->active ? "active" : "inactive", address->plugin, address->addr,
GNUNET_i2s (&address->peer));
}
unsigned int n_new;
int i;
struct RIL_Network *net;
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
- LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_delete() Delete %s%s %s address %p for peer '%s'\n",
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_delete() Delete %s%s %s address %s for peer '%s'\n",
session_only ? "session for " : "", address->active ? "active" : "inactive", address->plugin,
address->addr, GNUNET_i2s (&address->peer));
m_new = agent->m - RIL_FEATURES_ADDRESS_COUNT;
n_new = agent->n - 1;
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "first\n");
+
for (i = 0; i < agent->n; i++)
{
ril_cut_from_vector ((void **) &agent->W[i], sizeof(double),
- ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ ((RIL_FEATURES_NETWORK_COUNT) //TODO! replace, when adding more networks
+ (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
}
GNUNET_free(agent->W[RIL_ACTION_TYPE_NUM + address_index]);
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "second\n");
ril_cut_from_vector ((void **) &agent->W, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
1, agent->n);
//correct last action
agent->a_old = RIL_ACTION_INVALID;
}
//decrease old state vector and eligibility vector
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "third\n");
ril_cut_from_vector ((void **) &agent->s_old, sizeof(double),
- ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ ((RIL_FEATURES_NETWORK_COUNT) //TODO! replace when adding more networks
+ (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
ril_cut_from_vector ((void **) &agent->e, sizeof(double),
- ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ ((RIL_FEATURES_NETWORK_COUNT) //TODO! replace when adding more networks
+ (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
agent->m = m_new;
agent->n = n_new;
- LOG(GNUNET_ERROR_TYPE_DEBUG, "address was used: %s\n", address_was_used ? "yes" : "no");
-
if (address_was_used)
{
- net = address->solver_information;
- net->bw_in_assigned -= agent->bw_in;
- net->bw_out_assigned -= agent->bw_out;
-
if (NULL != agent->addresses_head) //if peer has an address left, use it
{
- LOG(GNUNET_ERROR_TYPE_DEBUG, "address left: %p\n",
- agent->addresses_head->address_naked->addr);
- //TODO? check if network/bandwidth update can be done more clever/elegant at different function
- envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, min_bw, min_bw,
+ envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, MIN_BW, MIN_BW,
GNUNET_NO);
- net = agent->addresses_head->address_naked->solver_information;
- net->bw_in_assigned -= min_bw;
- net->bw_out_assigned -= min_bw;
}
else
{
- LOG(GNUNET_ERROR_TYPE_DEBUG, "no address left => disconnect\n");
-
envi_set_active_suggestion (s, agent, NULL, 0, 0, GNUNET_NO);
}
}
- LOG(GNUNET_ERROR_TYPE_DEBUG, "Address deleted\n");
+ ril_step (solver);
}
/**
double rel_value)
{
LOG(GNUNET_ERROR_TYPE_DEBUG,
- "API_address_property_changed() Property '%s' for peer '%s' address %p changed "
+ "API_address_property_changed() Property '%s' for peer '%s' address %s changed "
"to %.2f \n", GNUNET_ATS_print_property_type (type), GNUNET_i2s (&address->peer),
address->addr, rel_value);
- /*
- * Nothing to do here, properties are considered in every reward calculation
- */
+
+ ril_step (solver);
}
/**
void
GAS_ril_address_inuse_changed (void *solver, struct ATS_Address *address, int in_use)
{
- /* Nothing to do here.
- * Possible TODO? Future Work: Use usage as state vector
+ /*
+ * TODO? Future Work: Potentially add usage variable to state vector
*/
LOG(GNUNET_ERROR_TYPE_DEBUG,
"API_address_inuse_changed() Usage for %s address of peer '%s' changed to %s\n",
{
struct GAS_RIL_Handle *s = solver;
struct RIL_Peer_Agent *agent;
- struct RIL_Network *net;
- uint32_t min_bw = ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__);
LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_change_network() Network type changed, moving "
"%s address of peer %s from '%s' to '%s'\n",
return;
}
- net = ril_get_network (s, current_network);
- net->bw_in_assigned -= agent->bw_in;
- net->bw_out_assigned -= agent->bw_out;
-
- net = ril_get_network (s, new_network);
- net->bw_in_assigned -= min_bw;
- net->bw_out_assigned -= min_bw;
- address->solver_information = net;
+ address->solver_information = ril_get_network(solver, new_network);
}
/**
/**
* Start a bulk operation
*
- * Since new calculations of the assignment are not triggered by a change of preferences, as it
- * happens in the proportional and the mlp solver, there is no need to block this solver.
- *
* @param solver the solver
*/
void
GAS_ril_bulk_start (void *solver)
{
- /* nop */
+ struct GAS_RIL_Handle *s = solver;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_start() lock: %d\n", s->bulk_lock+1);
+
+ s->bulk_lock++;
}
/**
* Bulk operation done
*
- * Since new calculations of the assignment are not triggered by a change of preferences, as it
- * happens in the proportional and the mlp solver, there is no need to block this solver.
- *
* @param solver the solver handle
*/
void
GAS_ril_bulk_stop (void *solver)
{
- /* nop */
+ struct GAS_RIL_Handle *s = solver;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_stop() lock: %d\n", s->bulk_lock-1);
+
+ if (s->bulk_lock < 1)
+ {
+ GNUNET_break(0);
+ return;
+ }
+ s->bulk_lock--;
+
+ if (0 < s->bulk_changes)
+ {
+ ril_step (solver);
+ s->bulk_changes = 0;
+ }
}
/**
*/
struct GAS_RIL_Handle *s = solver;
struct RIL_Peer_Agent *agent;
- struct RIL_Network *net;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_get_preferred_address()\n");
agent = ril_get_agent (s, peer, GNUNET_YES);
- agent->active = GNUNET_YES;
+ agent->is_active = GNUNET_YES;
+ envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, agent->bw_out, GNUNET_YES);
- if (agent->address_inuse)
- {
- net = agent->address_inuse->solver_information;
- net->bw_in_assigned += agent->bw_in;
- net->bw_out_assigned += agent->bw_out;
- }
- envi_set_active_suggestion (s, agent, agent->address_inuse, agent->bw_in, agent->bw_out,
- GNUNET_YES);
+ ril_try_unblock_agent(solver, agent, GNUNET_YES);
if (agent->address_inuse)
{
LOG(GNUNET_ERROR_TYPE_DEBUG,
- "API_get_preferred_address() Activated agent for peer '%s' with %s address %p\n",
+ "API_get_preferred_address() Activated agent for peer '%s' with %s address %s\n",
GNUNET_i2s (peer), agent->address_inuse->plugin, agent->address_inuse->addr);
}
else
{
struct GAS_RIL_Handle *s = solver;
struct RIL_Peer_Agent *agent;
- struct RIL_Network *net;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_stop_get_preferred_address()");
agent = ril_get_agent (s, peer, GNUNET_NO);
GNUNET_break(0);
return;
}
- if (GNUNET_NO == agent->active)
+ if (GNUNET_NO == agent->is_active)
{
GNUNET_break(0);
return;
}
- agent->active = GNUNET_NO;
- if (agent->address_inuse)
- {
- net = agent->address_inuse->solver_information;
- net->bw_in_assigned -= agent->bw_in;
- net->bw_out_assigned -= agent->bw_out;
- }
+ agent->is_active = GNUNET_NO;
+
envi_set_active_suggestion (s, agent, agent->address_inuse, agent->bw_in, agent->bw_out,
GNUNET_YES);
+ ril_step (s);
+
LOG(GNUNET_ERROR_TYPE_DEBUG,
"API_stop_get_preferred_address() Paused agent for peer '%s' with %s address\n",
GNUNET_i2s (peer), agent->address_inuse->plugin);
projects / oweals / gnunet.git / blobdiff