#define RIL_ACTION_INVALID -1
#define RIL_FEATURES_ADDRESS_COUNT (3 + GNUNET_ATS_QualityPropertiesCount)
#define RIL_FEATURES_NETWORK_COUNT 4
+#define RIL_INTERVAL_EXPONENT 10
-#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_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 0.7
#define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.4
#define RIL_DEFAULT_TRACE_DECAY 0.6
-#define RIL_EXPLORE_RATIO 0.1
+#define RIL_DEFAULT_EXPLORE_RATIO 0.1
/**
* ATS reinforcement learning solver
/**
* Learning discount factor in the TD-update
*/
- float gamma;
+ double beta;
/**
* Gradient-descent step-size
*/
- float alpha;
+ double alpha;
/**
* Trace-decay factor for eligibility traces
*/
- float lambda;
+ double lambda;
+
+ /**
+ *
+ */
+ double explore_ratio;
+
+ /**
+ * 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; //TODO? rename into something better. It reflects the state whether get_preferred_address() has been called for the according peer or not
+ int is_active;
/**
* Number of performed time-steps
unsigned long long bw_out_assigned;
};
-struct RIL_Callbacks
+/**
+ * A handle for the reinforcement learning solver
+ */
+struct GAS_RIL_Handle
{
/**
- * Bandwidth changed callback
+ * The solver-plugin environment of the solver-plugin API
*/
- GAS_bandwidth_changed_cb bw_changed;
-
- /**
- * Bandwidth changed callback cls
- */
- void *bw_changed_cls;
+ struct GNUNET_ATS_PluginEnvironment *plugin_envi;
/**
- * ATS function to get preferences for a peer
+ * Statistics handle
*/
- GAS_get_preferences get_preferences;
+ struct GNUNET_STATISTICS_Handle *stats;
/**
- * Closure for ATS function to get preferences
+ * Number of performed steps
*/
- void *get_preferences_cls;
+ unsigned long long step_count;
/**
- * ATS function to get properties of an address
+ * Timestamp for the last time-step
*/
- GAS_get_properties get_properties;
+ struct GNUNET_TIME_Absolute step_time_last;
/**
- * Closure for ATS function to get properties
- */
- void *get_properties_cls;
-};
-
-/**
- * A handle for the reinforcement learning solver
- */
-struct GAS_RIL_Handle
-{
- /**
- *
+ * Task identifier of the next time-step to be executed
*/
- struct GNUNET_ATS_PluginEnvironment *plugin_envi;
+ GNUNET_SCHEDULER_TaskIdentifier step_next_task_id;
/**
- * Statistics handle
+ * Whether a step is already scheduled
*/
- struct GNUNET_STATISTICS_Handle *stats;
+ int task_pending;
/**
- * Hashmap containing all valid addresses
+ * Variable discount factor, dependent on time between steps
*/
- const struct GNUNET_CONTAINER_MultiPeerMap *addresses;
+ double discount_variable;
/**
- * Callbacks for the solver
+ * Integrated variable discount factor, dependent on time between steps
*/
- struct RIL_Callbacks *callbacks;
+ double discount_integrated;
/**
- * Number of performed time-steps
+ * Lock for bulk operations
*/
- unsigned long long step_count;
+ int bulk_lock;
/**
- * Interval time between steps in milliseconds //TODO? Future Work: Heterogeneous stepping among agents
+ * Number of changes during a lock
*/
- struct GNUNET_TIME_Relative step_time;
-
- /**
- * Task identifier of the next time-step to be executed
- */
- GNUNET_SCHEDULER_TaskIdentifier next_step;
+ int bulk_changes;
/**
* Learning parameters
/**
* Estimate the current action-value for state s and action a
+ *
* @param agent agent performing the estimation
* @param state s
* @param action a
/**
* Decide whether to do exploration (i.e. taking a new action) or exploitation (i.e. taking the
* currently estimated best action) in the current step
+ *
* @param agent agent performing the step
* @return yes, if exploring
*/
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;
}
return GNUNET_NO;
}
-/**
- * Get the index of the address in the agent's list.
- * @param agent agent handle
- * @param address address handle
- * @return the index, starting with zero
- */
+ /**
+ * Get the index of the address in the agent's list.
+ *
+ * @param agent agent handle
+ * @param address address handle
+ * @return the index, starting with zero
+ */
static int
agent_address_get_index (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
{
/**
* Gets the wrapped address from the agent's list
+ *
* @param agent agent handle
* @param address address handle
* @return wrapped address
/**
* Gets the action, with the maximal estimated Q-value (i.e. the one currently estimated to bring the
* most reward in the future)
+ *
* @param agent agent performing the calculation
* @param state the state from which to take the action
* @return the action promising most future reward
/**
* Gets any action, to explore the action space from that state
+ *
* @param agent agent performing the calculation
* @param state the state from which to take the action
* @return any action
/**
* Updates the weights (i.e. coefficients) of the weight vector in matrix W for action a
+ *
* @param agent the agent performing the update
* @param reward the reward received for the last action
* @param s_next the new state, the last step got the agent into
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->discount_integrated * reward; //reward
+ delta += agent->envi->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
for (i = 0; i < agent->m; i++)
{
- theta[i] += agent->envi->parameters.alpha * delta * (agent->e)[i];
+ theta[i] += agent->envi->parameters.alpha * delta * 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
- * @param mod
+ *
+ * @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)
{
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++)
{
e[i] = 1;
break;
case RIL_E_SET:
- e[i] = gamma * lambda;
+ e[i] *= agent->envi->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.
+ * notify ATS of its new decision
+ *
* @param solver solver handle
* @param agent agent handle
* @param new_address the address which is to be used
{
int notify = GNUNET_NO;
- LOG (GNUNET_ERROR_TYPE_DEBUG, "set_active_suggestion()\n");
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "set_active_suggestion()\n");
//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;
}
//bw change
}
}
- if (notify && agent->active && (GNUNET_NO == silent))
+ if (notify && agent->is_active && (GNUNET_NO == silent))
{
if (new_address)
{
- solver->callbacks->bw_changed (solver->callbacks->bw_changed_cls, new_address);
+ solver->plugin_envi->bandwidth_changed_cb (solver->plugin_envi->bw_changed_cb_cls,
+ new_address);
}
- else
+ else if (agent->address_inuse)
{
- GNUNET_assert (0 == ntohl(agent->address_inuse->assigned_bw_in.value__));
- GNUNET_assert (0 == ntohl(agent->address_inuse->assigned_bw_out.value__));
+ 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->callbacks->bw_changed (solver->callbacks->bw_changed_cls, agent->address_inuse);
+ solver->plugin_envi->bandwidth_changed_cb (solver->plugin_envi->bw_changed_cb_cls,
+ agent->address_inuse);
}
}
agent->address_inuse = new_address;
/**
* 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 *
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->callbacks->get_properties (solver->callbacks->get_properties_cls, cur_address->address_naked);
+ 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];
/**
* 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)
+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'
/**
* Get the index of the quality-property in question
+ *
* @param type the quality property type
* @return the index
*/
return GNUNET_SYSERR;
}
-/**
- * Gets the reward of the last performed step
- * @param solver solver handle
- * @return the reward
- */
static double
-envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
+envi_reward_global (struct GAS_RIL_Handle *solver)
{
- /*
- * Match the preferences of the peer with the current assignment.
- */
+ int i;
+ unsigned int in_available = 0;
+ unsigned int out_available = 0;
+ unsigned int in_assigned = 0;
+ unsigned int out_assigned = 0;
+ double ratio_in;
+ double ratio_out;
+
+ for (i = 0; i < solver->networks_count; i++)
+ {
+ in_available += solver->network_entries[i].bw_in_available;
+ in_assigned += solver->network_entries[i].bw_in_assigned;
+ out_available += solver->network_entries[i].bw_out_available;
+ out_assigned += solver->network_entries[i].bw_out_assigned;
+ }
+
+ ratio_in = ((double) in_assigned) / ((double) in_available);
+ ratio_out = ((double) out_assigned) / ((double) out_available);
+
+ return ((ratio_in + ratio_out) * 0.5) + 1;
+}
+
+static double
+envi_reward_local (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
+{
+ //TODO! add utilization
+
const double *preferences;
const double *properties;
+ int prop_index;
double pref_match = 0;
double bw_norm;
- struct RIL_Network *net;
- int prop_index;
- preferences = solver->callbacks->get_preferences (solver->callbacks->get_preferences_cls, &agent->peer);
- properties = solver->callbacks->get_properties (solver->callbacks->get_properties_cls,
+ 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] * properties[prop_index];
+
+ //preference matching from latency and bandwidth
+ prop_index = ril_find_property_index (GNUNET_ATS_QUALITY_NET_DELAY);
+ pref_match += 1 - (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) 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 += 1 - (preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] * bw_norm);
+ return pref_match * 0.5;
+}
+
+/**
+ * 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 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)
+{
+ struct RIL_Network *net;
+ double reward = 0;
+
+ //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)
+ || (net->bw_out_assigned > net->bw_out_available))
{
return -1;
}
- return pref_match;
+ reward += envi_reward_global (solver);
+ reward += envi_reward_local (solver, agent);
+ return reward * 0.5;
}
/**
* Doubles the bandwidth for the active address
+ *
* @param solver solver handle
* @param agent agent handle
* @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise the outbound bandwidth
/**
* Cuts the bandwidth for the active address in half. The least amount of bandwidth suggested, is
* the minimum bandwidth for a peer, in order to not invoke a disconnect.
+ *
* @param solver solver handle
* @param agent agent handle
* @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
new_bw = agent->bw_in / 2;
if (new_bw < min_bw)
new_bw = min_bw;
- envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out, GNUNET_NO);
+ 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;
- envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw, GNUNET_NO);
+ envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
+ GNUNET_NO);
}
}
/**
* Increases the bandwidth by 5 times the minimum bandwidth for the active address.
+ *
* @param solver solver handle
* @param agent agent handle
* @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
/**
* Decreases the bandwidth by 5 times the minimum bandwidth for the active address. The least amount
* of bandwidth suggested, is the minimum bandwidth for a peer, in order to not invoke a disconnect.
+ *
* @param solver solver handle
* @param agent agent handle
* @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
new_bw = agent->bw_in - (5 * min_bw);
if (new_bw < min_bw)
new_bw = min_bw;
- envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out, GNUNET_NO);
+ 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;
- envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw, GNUNET_NO);
+ envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
+ GNUNET_NO);
}
}
/**
* Switches to the address given by its index
+ *
* @param solver solver handle
* @param agent agent handle
* @param address_index index of the address as it is saved in the agent's list, starting with zero
{
if (i == address_index)
{
- envi_set_active_suggestion (solver, agent, cur->address_naked, agent->bw_in, agent->bw_out, GNUNET_NO);
+ envi_set_active_suggestion (solver, agent, cur->address_naked, agent->bw_in, agent->bw_out,
+ GNUNET_NO);
return;
}
/**
* 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)
{
- unsigned int address_index;
+ int address_index;
switch (action)
{
* after having done the last action a_old. It observes the new state s_next and the reward
* received. Then the coefficient update is done according to the SARSA or Q-learning method. The
* next action is put into effect.
+ *
* @param agent the agent performing the step
*/
static void
agent->step_count += 1;
}
+static int
+ril_step (struct GAS_RIL_Handle *solver);
+
/**
- * Cycles through all agents and lets the active ones do a step. Schedules the next step.
- * @param solver the solver handle
- * @param tc task context for the scheduler
+ * Task for the scheduler, which performs one step and lets the solver know that
+ * no further step is scheduled.
+ *
+ * @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->task_pending = GNUNET_NO;
+ 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 with an active address
+ {
+ 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
+}
+
+/**
+ * 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;
+
+ if (solver->task_pending)
+ {
+ GNUNET_SCHEDULER_cancel (solver->step_next_task_id);
+ }
+
+ 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;
+
+ //LOG (GNUNET_ERROR_TYPE_INFO, "used = %f, min = %f, max = %f\n", used_ratio, (double)solver->parameters.step_time_min.rel_value_us, (double)solver->parameters.step_time_max.rel_value_us);
+ //LOG (GNUNET_ERROR_TYPE_INFO, "factor = %f, offset = %f, y = %f\n", factor, offset, y);
+
+ 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 int) y);
+
+ solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (time_next, &ril_step_scheduler_task,
+ solver);
+ solver->task_pending = GNUNET_YES;
+}
+
+/**
+ * Triggers one step per agent
+ * @param solver
+ */
+static int
+ril_step (struct GAS_RIL_Handle *solver)
+{
struct RIL_Peer_Agent *cur;
+ struct GNUNET_TIME_Absolute time_now;
+ struct GNUNET_TIME_Relative time_delta;
+ double tau;
+
+ if (GNUNET_YES == solver->bulk_lock)
+ {
+ solver->bulk_changes++;
+ return GNUNET_NO;
+ }
+
+ 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 ();
+ }
+
+ //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);
+ tau = ((double) time_delta.rel_value_us)
+ / ((double) solver->parameters.step_time_min.rel_value_us);
+ solver->step_time_last = time_now;
+
+ //calculate reward discounts (once per step for all agents)
+ solver->discount_variable = pow (M_E, ((-1.) * ((double) solver->parameters.beta) * tau));
+ solver->discount_integrated = (1 - solver->discount_variable)
+ / ((double) solver->parameters.beta);
+
+ //trigger one step per active agent
for (cur = solver->agents_head; NULL != cur; cur = cur->next)
{
- if (cur->active && cur->address_inuse)
+ if (cur->is_active && cur->address_inuse)
{
agent_step (cur);
}
}
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);
+
+ return GNUNET_YES;
}
/**
* Initialize an agent without addresses and its knowledge base
+ *
* @param s ril solver
* @param peer the one in question
* @return handle to the new agent
agent->envi = solver;
agent->peer = *peer;
agent->step_count = 0;
- agent->active = GNUNET_NO;
+ agent->is_active = GNUNET_NO;
agent->n = RIL_ACTION_TYPE_NUM;
agent->m = solver->networks_count * RIL_FEATURES_NETWORK_COUNT;
- agent->W = (double **) GNUNET_malloc (sizeof (double) * agent->n);
+ 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);
/**
* 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)
for (cur = solver->agents_head; NULL != cur; cur = cur->next)
{
- if (0 == memcmp (peer,
- &cur->peer,
- sizeof (struct GNUNET_PeerIdentity)))
+ if (0 == memcmp (peer, &cur->peer, sizeof(struct GNUNET_PeerIdentity)))
{
return cur;
}
/**
* Determine whether at least the minimum bandwidth is set for the network. Otherwise the network is
* considered inactive and not used. Addresses in an inactive network are ignored.
+ *
* @param solver solver handle
* @param network the network type
- * @return
+ * @return whether or not the network is considered active
*/
static int
ril_network_is_active (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
* Cuts a slice out of a vector of elements. This is used to decrease the size of the matrix storing
* the reward function approximation. It copies the memory, which is not cut, to the new vector,
* frees the memory of the old vector, and redirects the pointer to the new one.
+ *
* @param old pointer to the pointer to the first element of the vector
* @param element_size byte size of the vector elements
* @param hole_start the first element to cut out
*/
/**
- * Changes the preferences for a peer in the problem
+ * Change relative preference for quality in solver
*
* @param solver the solver handle
* @param peer the peer to change the preference for
* @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);
}
-//TODO doxygen
+/**
+ * Entry point for the plugin
+ *
+ * @param cls pointer to the 'struct GNUNET_ATS_PluginEnvironment'
+ */
void *
libgnunet_plugin_ats_ril_init (void *cls)
{
struct GNUNET_ATS_PluginEnvironment *env = cls;
- struct GAS_RIL_Handle *solver = GNUNET_new (struct GAS_RIL_Handle);;
+ 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->parameters.step_time_min = RIL_DEFAULT_STEP_TIME_MIN;
+ }
+ if (GNUNET_OK
+ != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MAX",
+ &solver->parameters.step_time_max))
{
- solver->step_time = RIL_DEFAULT_STEP_TIME;
+ solver->parameters.step_time_max = RIL_DEFAULT_STEP_TIME_MAX;
}
- 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_string (env->cfg, "ats", "RIL_ALGORITHM", &string))
{
- solver->parameters.algorithm = RIL_ALGO_SARSA;
+ 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_size (env->cfg, "ats", "RIL_GRADIENT_STEP_SIZE", &tmp))
+ if (GNUNET_OK
+ == 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;
+ }
env->sf.s_add = &GAS_ril_address_add;
env->sf.s_address_update_property = &GAS_ril_address_property_changed;
env->sf.s_bulk_stop = &GAS_ril_bulk_stop;
solver->plugin_envi = env;
- solver->stats = (struct GNUNET_STATISTICS_Handle *) env->stats;
- solver->callbacks = GNUNET_malloc (sizeof (struct RIL_Callbacks));
- solver->callbacks->bw_changed = env->bandwidth_changed_cb;
- solver->callbacks->bw_changed_cls = env->bw_changed_cb_cls;
- solver->callbacks->get_preferences = env->get_preferences;
- solver->callbacks->get_preferences_cls = env->get_preference_cls;
- solver->callbacks->get_properties = env->get_property;
- solver->callbacks->get_properties_cls = env->get_property_cls;
solver->networks_count = env->network_count;
solver->network_entries = GNUNET_malloc (env->network_count * sizeof (struct RIL_Network));
- solver->addresses = env->addresses;
solver->step_count = 0;
for (c = 0; c < env->network_count; c++)
cur->bw_out_assigned = 0;
}
- solver->next_step = GNUNET_SCHEDULER_add_delayed (
+ solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (
GNUNET_TIME_relative_multiply (GNUNET_TIME_relative_get_millisecond_ (), 1000),
- &ril_periodic_step, solver);
+ &ril_step_scheduler_task, solver);
+ solver->task_pending = GNUNET_YES;
return solver;
}
-//TODO doxygen
+/**
+ * Exit point for the plugin
+ *
+ * @param cls the solver handle
+ */
void *
libgnunet_plugin_ats_ril_done (void *cls)
{
cur_agent = next_agent;
}
- GNUNET_SCHEDULER_cancel (s->next_step);
- GNUNET_free(s->callbacks);
+ if (s->task_pending)
+ {
+ GNUNET_SCHEDULER_cancel (s->step_next_task_id);
+ }
GNUNET_free(s->network_entries);
GNUNET_free(s);
- return NULL;
+ return NULL ;
}
/**
- * Add a single address within a network to the solver
+ * Add a new address for a peer to the solver
+ *
+ * The address is already contained in the addresses hashmap!
*
* @param solver the solver Handle
* @param address the address to add
//increase size of old state vector
agent->m = m_old;
- GNUNET_array_grow(agent->s_old, agent->m, m_new); //TODO random instead of zero-initialization of state features
+ GNUNET_array_grow(agent->s_old, agent->m, m_new);
agent->m = m_old;
GNUNET_array_grow(agent->e, agent->m, m_new);
envi_set_active_suggestion (s, agent, address, min_bw, min_bw, GNUNET_NO);
}
+ ril_step (s);
+
LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %p for peer '%s'\n",
- address->active ? "active" : "inactive",
- address->plugin, address->addr, GNUNET_i2s (&address->peer));
+ address->active ? "active" : "inactive", address->plugin, address->addr,
+ GNUNET_i2s (&address->peer));
}
/**
- * Remove an address from the solver
+ * Delete an address in the solver
+ *
+ * The address is not contained in the address hashmap anymore!
*
* @param solver the solver handle
* @param address the address to remove
for (i = 0; i < agent->n; i++)
{
ril_cut_from_vector ((void **) &agent->W[i], sizeof(double),
- ((s->networks_count * RIL_FEATURES_NETWORK_COUNT) + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
+ ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
}
- GNUNET_free (agent->W[RIL_ACTION_TYPE_NUM + address_index]);
+ GNUNET_free(agent->W[RIL_ACTION_TYPE_NUM + address_index]);
ril_cut_from_vector ((void **) &agent->W, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
1, agent->n);
//correct last action
}
//decrease old state vector and eligibility vector
ril_cut_from_vector ((void **) &agent->s_old, sizeof(double),
- ((s->networks_count * RIL_FEATURES_NETWORK_COUNT) + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
+ ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ + (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) + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
+ ((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
+ + (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");
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "address was used: %s\n", address_was_used ? "yes" : "no");
if (address_was_used)
{
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);
+ 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, GNUNET_NO);
+ 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");
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "no address left => disconnect\n");
envi_set_active_suggestion (s, agent, NULL, 0, 0, GNUNET_NO);
}
}
+ ril_step (solver);
+
LOG(GNUNET_ERROR_TYPE_DEBUG, "Address deleted\n");
}
/**
- * Transport properties for this address have changed
+ * Update the properties of an address in the solver
*
* @param solver solver handle
* @param address the address
"API_address_property_changed() Property '%s' for peer '%s' address %p 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);
}
/**
- * Transport session for this address has changed
+ * Update the session of an address in the solver
*
* NOTE: values in addresses are already updated
*
}
/**
- * Usage for this address has changed
+ * Notify the solver that an address is (not) actively used by transport
+ * to communicate with a remote peer
*
* NOTE: values in addresses are already updated
*
void
GAS_ril_address_inuse_changed (void *solver, struct ATS_Address *address, int in_use)
{
- /* Nothing to do here */
+ /* Nothing to do here.
+ * Possible 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",
address->plugin, GNUNET_i2s (&address->peer), (GNUNET_YES == in_use) ? "USED" : "UNUSED");
}
/**
- * Network scope for this address has changed
+ * Notify solver that the network an address is located in has changed
*
* NOTE: values in addresses are already updated
*
}
/**
- * Get application feedback for a peer
+ * Give feedback about the current assignment
*
* @param solver the solver handle
* @param application the application
enum GNUNET_ATS_PreferenceKind kind,
double score)
{
- //TODO! collect feedback
LOG(GNUNET_ERROR_TYPE_DEBUG,
"API_address_preference_feedback() Peer '%s' got a feedback of %+.3f from application %s for "
"preference %s for %d seconds\n", GNUNET_i2s (peer), "UNKNOWN",
void
GAS_ril_bulk_start (void *solver)
{
- /*
- * 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.
- */
+ struct GAS_RIL_Handle *s = solver;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_start() Locking solver for bulk operation ...\n");
+
+ s->bulk_lock++;
}
/**
* Bulk operation done
+ *
+ * @param solver the solver handle
*/
void
GAS_ril_bulk_stop (void *solver)
{
- /*
- * 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.
- */
+ struct GAS_RIL_Handle *s = solver;
+
+ LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_stop() Releasing solver from bulk operation ...\n");
+
+ if (s->bulk_lock < 1)
+ {
+ GNUNET_break(0);
+ return;
+ }
+ s->bulk_lock--;
+
+ if (0 < s->bulk_changes)
+ {
+ ril_step (solver);
+ s->bulk_changes = 0;
+ }
}
/**
- * Get the preferred address for a specific peer
+ * Tell solver to notify ATS if the address to use changes for a specific
+ * peer using the bandwidth changed callback
+ *
+ * The solver must only notify about changes for peers with pending address
+ * requests!
*
* @param solver the solver handle
* @param peer the identity of the peer
*/
struct GAS_RIL_Handle *s = solver;
struct RIL_Peer_Agent *agent;
+ struct RIL_Network *net;
agent = ril_get_agent (s, peer, GNUNET_YES);
- agent->active = GNUNET_YES;
+ agent->is_active = GNUNET_YES;
- envi_set_active_suggestion(s, 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);
if (agent->address_inuse)
{
else
{
LOG(GNUNET_ERROR_TYPE_DEBUG,
- "API_get_preferred_address() Activated agent for peer '%s', but no address available\n", GNUNET_i2s (peer));
+ "API_get_preferred_address() Activated agent for peer '%s', but no address available\n",
+ GNUNET_i2s (peer));
}
+ ril_step (s);
+
return agent->address_inuse;
}
/**
- * Stop notifying about address and bandwidth changes for this peer
+ * Tell solver stop notifying ATS about changes for this peers
+ *
+ * The solver must only notify about changes for peers with pending address
+ * requests!
*
* @param solver the solver handle
* @param peer the peer
{
struct GAS_RIL_Handle *s = solver;
struct RIL_Peer_Agent *agent;
+ struct RIL_Network *net;
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;
- envi_set_active_suggestion(s, agent, agent->address_inuse, agent->bw_in, agent->bw_out, GNUNET_YES);
+ agent->is_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;
+ }
+ 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",