2 This file is part of GNUnet.
3 (C) 2011 Christian Grothoff (and other contributing authors)
5 GNUnet is free software; you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published
7 by the Free Software Foundation; either version 3, or (at your
8 option) any later version.
10 GNUnet is distributed in the hope that it will be useful, but
11 WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with GNUnet; see the file COPYING. If not, write to the
17 Free Software Foundation, Inc., 59 Temple Place - Suite 330,
18 Boston, MA 02111-1307, USA.
22 * @file ats/plugin_ats_ril.c
23 * @brief ATS reinforcement learning solver
24 * @author Fabian Oehlmann
25 * @author Matthias Wachs
27 #include "plugin_ats_ril.h"
29 #define LOG(kind,...) GNUNET_log_from (kind, "ats-ril",__VA_ARGS__)
31 #define MIN_BW ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__)
33 #define RIL_ACTION_INVALID -1
34 #define RIL_FEATURES_ADDRESS_COUNT (0)// + GNUNET_ATS_QualityPropertiesCount)
35 #define RIL_FEATURES_NETWORK_COUNT 2
36 #define RIL_FEATURES_INIT_COUNT 1 + RIL_FEATURES_NETWORK_COUNT // + GNUNET_ATS_PreferenceCount
37 #define RIL_INTERVAL_EXPONENT 10
38 #define RIL_UTILITY_MAX (double) GNUNET_ATS_MaxBandwidth
40 #define RIL_DEFAULT_STEP_TIME_MIN GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 500)
41 #define RIL_DEFAULT_STEP_TIME_MAX GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 3000)
42 #define RIL_DEFAULT_ALGORITHM RIL_ALGO_SARSA
43 #define RIL_DEFAULT_DISCOUNT_BETA 1.0
44 #define RIL_DEFAULT_DISCOUNT_GAMMA 0.5
45 #define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.1
46 #define RIL_DEFAULT_TRACE_DECAY 0.5
47 #define RIL_DEFAULT_EXPLORE_RATIO 0.1
48 #define RIL_DEFAULT_GLOBAL_REWARD_SHARE 0.5
50 #define RIL_INC_DEC_STEP_SIZE 1
53 * ATS reinforcement learning solver
59 * The actions, how an agent can manipulate the current assignment. I.e. how the bandwidth can be
60 * changed for the currently chosen address. Not depicted in the enum are the actions of switching
61 * to a particular address. The action of switching to address with index i is depicted by the
62 * number (RIL_ACTION_TYPE_NUM + i).
66 RIL_ACTION_NOTHING = -1,
67 RIL_ACTION_BW_IN_DBL = -2, //TODO! put actions back
68 RIL_ACTION_BW_IN_HLV = -3,
69 RIL_ACTION_BW_IN_INC = 0,
70 RIL_ACTION_BW_IN_DEC = 1,
71 RIL_ACTION_BW_OUT_DBL = -4,
72 RIL_ACTION_BW_OUT_HLV = -5,
73 RIL_ACTION_BW_OUT_INC = -6,
74 RIL_ACTION_BW_OUT_DEC = -7,
75 RIL_ACTION_TYPE_NUM = 1
84 enum RIL_E_Modification
93 * Global learning parameters
95 struct RIL_Learning_Parameters
98 * The TD-algorithm to use
100 enum RIL_Algorithm algorithm;
103 * Gradient-descent step-size
108 * Learning discount variable in the TD-update for semi-MDPs
113 * Learning discount factor in the TD-update for MDPs
118 * Trace-decay factor for eligibility traces
123 * Ratio, with what probability an agent should explore in the e-greed policy
125 double explore_ratio;
128 * How big the share of the global part of the reward signal is
130 double reward_global_share;
133 * Minimal interval time between steps in milliseconds
135 struct GNUNET_TIME_Relative step_time_min;
138 * Maximum interval time between steps in milliseconds
140 struct GNUNET_TIME_Relative step_time_max;
144 * Wrapper for addresses to store them in agent's linked list
146 struct RIL_Address_Wrapped
151 struct RIL_Address_Wrapped *next;
156 struct RIL_Address_Wrapped *prev;
161 struct ATS_Address *address_naked;
164 struct RIL_Peer_Agent
167 * Next agent in solver's linked list
169 struct RIL_Peer_Agent *next;
172 * Previous agent in solver's linked list
174 struct RIL_Peer_Agent *prev;
179 struct GAS_RIL_Handle *envi;
184 struct GNUNET_PeerIdentity peer;
187 * Whether the agent is active or not
192 * Number of performed time-steps
194 unsigned long long step_count;
197 * Experience matrix W
202 * Number of rows of W / Number of state-vector features
207 * Number of columns of W / Number of actions
212 * Last perceived state feature vector
222 * Eligibility trace vector
229 struct ATS_Address * address_inuse;
232 * Head of addresses DLL
234 struct RIL_Address_Wrapped * addresses_head;
237 * Tail of addresses DLL
239 struct RIL_Address_Wrapped * addresses_tail;
242 * Inbound bandwidth assigned by the agent
244 unsigned long long bw_in;
247 * Outbound bandwidth assigned by the agent
249 unsigned long long bw_out;
252 * Flag whether a suggestion has to be issued
254 int suggestion_issue;
257 * The address which has to be issued
259 struct ATS_Address * suggestion_address;
267 enum GNUNET_ATS_Network_Type type;
270 * Total available inbound bandwidth
272 unsigned long long bw_in_available;
275 * Bandwidth inbound assigned in network after last step
277 unsigned long long bw_in_assigned;
280 * Total available outbound bandwidth
282 unsigned long long bw_out_available;
285 * * Bandwidth outbound assigned in network after last step
287 unsigned long long bw_out_assigned;
291 * A handle for the reinforcement learning solver
293 struct GAS_RIL_Handle
296 * The solver-plugin environment of the solver-plugin API
298 struct GNUNET_ATS_PluginEnvironment *plugin_envi;
303 struct GNUNET_STATISTICS_Handle *stats;
306 * Number of performed steps
308 unsigned long long step_count;
311 * Timestamp for the last time-step
313 struct GNUNET_TIME_Absolute step_time_last;
316 * Task identifier of the next time-step to be executed
318 GNUNET_SCHEDULER_TaskIdentifier step_next_task_id;
321 * Variable discount factor, dependent on time between steps
323 double global_discount_variable;
326 * Integrated variable discount factor, dependent on time between steps
328 double global_discount_integrated;
331 * State vector for networks for the current step
333 double *global_state_networks;
336 * Lock for bulk operations
341 * Number of changes during a lock
346 * Learning parameters
348 struct RIL_Learning_Parameters parameters;
351 * Array of networks with global assignment state
353 struct RIL_Network * network_entries;
358 unsigned int networks_count;
361 * List of active peer-agents
363 struct RIL_Peer_Agent * agents_head;
364 struct RIL_Peer_Agent * agents_tail;
372 * Simulate steps, i.e. schedule steps immediately
374 unsigned long long simulate;
379 * ---------------------------
383 ril_count_agents(struct GAS_RIL_Handle * solver);
386 agent_get_utility (struct RIL_Peer_Agent *agent)
388 return (double) agent->bw_in;
392 * Estimate the current action-value for state s and action a
394 * @param agent agent performing the estimation
397 * @return estimation value
400 agent_estimate_q (struct RIL_Peer_Agent *agent, double *state, int action)
405 for (i = 0; i < agent->m; i++)
407 result += state[i] * agent->W[action][i];
410 GNUNET_assert(!isnan(result));
414 return isinf(result) * UINT32_MAX; //TODO! prevent crash when learning diverges
421 * Decide whether to do exploration (i.e. taking a new action) or exploitation (i.e. taking the
422 * currently estimated best action) in the current step
424 * @param agent agent performing the step
425 * @return yes, if exploring
428 agent_decide_exploration (struct RIL_Peer_Agent *agent)
430 //TODO? Future Work: Improve exploration/exploitation trade-off by different mechanisms than e-greedy
431 double r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
432 UINT32_MAX) / (double) UINT32_MAX;
434 if (r < agent->envi->parameters.explore_ratio)
443 * Get the index of the address in the agent's list.
445 * @param agent agent handle
446 * @param address address handle
447 * @return the index, starting with zero
450 agent_address_get_index (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
453 struct RIL_Address_Wrapped *cur;
456 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
459 if (cur->address_naked == address)
467 * Gets the wrapped address from the agent's list
469 * @param agent agent handle
470 * @param address address handle
471 * @return wrapped address
473 static struct RIL_Address_Wrapped *
474 agent_address_get (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
476 struct RIL_Address_Wrapped *cur;
478 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
479 if (cur->address_naked == address)
486 * Gets the action, with the maximal estimated Q-value (i.e. the one currently estimated to bring the
487 * most reward in the future)
489 * @param agent agent performing the calculation
490 * @param state the state from which to take the action
491 * @return the action promising most future reward
494 agent_get_action_best (struct RIL_Peer_Agent *agent, double *state)
497 int max_i = RIL_ACTION_INVALID;
499 double max_q = -DBL_MAX;
501 for (i = 0; i < agent->n; i++)
503 cur_q = agent_estimate_q (agent, state, i);
511 GNUNET_assert(RIL_ACTION_INVALID != max_i);
518 * Gets any action, to explore the action space from that state
520 * @param agent agent performing the calculation
521 * @param state the state from which to take the action
525 agent_get_action_explore (struct RIL_Peer_Agent *agent, double *state)
527 // TODO?: Future Work: Choose the action for exploration, which has been explored the least in this state
528 return GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK, agent->n);
533 * Updates the weights (i.e. coefficients) of the weight vector in matrix W for action a
535 * @param agent the agent performing the update
536 * @param reward the reward received for the last action
537 * @param s_next the new state, the last step got the agent into
538 * @param a_prime the new
541 agent_update_weights (struct RIL_Peer_Agent *agent, double reward, double *s_next, int a_prime)
545 double *theta = agent->W[agent->a_old];
547 delta = agent->envi->global_discount_integrated * reward; //reward
548 delta += agent->envi->global_discount_variable * agent_estimate_q (agent, s_next, a_prime); //discounted future value
549 delta -= agent_estimate_q (agent, agent->s_old, agent->a_old); //one step
551 // 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",
552 // agent->step_count,
553 // agent_estimate_q (agent, agent->s_old, agent->a_old),
554 // agent->envi->parameters.alpha,
556 // agent->envi->global_discount_variable,
557 // agent_estimate_q (agent, s_next, a_prime),
560 for (i = 0; i < agent->m; i++)
562 // LOG(GNUNET_ERROR_TYPE_INFO, "alpha = %f delta = %f e[%d] = %f\n",
563 // agent->envi->parameters.alpha,
567 theta[i] += agent->envi->parameters.alpha * delta * agent->s_old[i];// * agent->e[i];
573 * Changes the eligibility trace vector e in various manners:
574 * #RIL_E_ACCUMULATE - adds @a f to each component as in accumulating eligibility traces
575 * #RIL_E_REPLACE - resets each component to @a f as in replacing traces
576 * #RIL_E_SET - multiplies e with discount factor and lambda as in the update rule
577 * #RIL_E_ZERO - sets e to 0 as in Watkin's Q-learning algorithm when exploring and when initializing
579 * @param agent the agent handle
580 * @param mod the kind of modification
581 * @param f how much to change
584 agent_modify_eligibility (struct RIL_Peer_Agent *agent,
585 enum RIL_E_Modification mod,
589 double *e = agent->e;
591 for (i = 0; i < agent->m; i++)
595 case RIL_E_ACCUMULATE:
602 e[i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
613 ril_inform (struct GAS_RIL_Handle *solver,
614 enum GAS_Solver_Operation op,
615 enum GAS_Solver_Status stat)
617 if (NULL != solver->plugin_envi->info_cb)
618 solver->plugin_envi->info_cb (solver->plugin_envi->info_cb_cls, op, stat, GAS_INFO_NONE);
623 * Changes the active assignment suggestion of the handler and invokes the bw_changed callback to
624 * notify ATS of its new decision
626 * @param solver solver handle
627 * @param agent agent handle
628 * @param new_address the address which is to be used
629 * @param new_bw_in the new amount of inbound bandwidth set for this address
630 * @param new_bw_out the new amount of outbound bandwidth set for this address
631 * @param silent disables invocation of the bw_changed callback, if GNUNET_YES
634 envi_set_active_suggestion (struct GAS_RIL_Handle *solver,
635 struct RIL_Peer_Agent *agent,
636 struct ATS_Address *new_address,
637 unsigned long long new_bw_in,
638 unsigned long long new_bw_out,
641 int notify = GNUNET_NO;
643 LOG(GNUNET_ERROR_TYPE_DEBUG, " set_active_suggestion() for peer '%s'\n", GNUNET_i2s (&agent->peer));
646 if (agent->address_inuse != new_address)
648 if (NULL != agent->address_inuse)
650 agent->address_inuse->active = GNUNET_NO;
651 agent->address_inuse->assigned_bw_in.value__ = htonl (0);
652 agent->address_inuse->assigned_bw_out.value__ = htonl (0);
654 if (NULL != new_address)
656 LOG(GNUNET_ERROR_TYPE_DEBUG, " set address active: %s\n", agent->is_active ? "yes" : "no");
657 new_address->active = agent->is_active;
658 new_address->assigned_bw_in.value__ = htonl (agent->bw_in);
659 new_address->assigned_bw_out.value__ = htonl (agent->bw_out);
661 notify |= GNUNET_YES;
667 if (new_address->active != agent->is_active)
669 new_address->active = agent->is_active;
670 notify |= GNUNET_YES;
674 if (agent->bw_in != new_bw_in)
676 agent->bw_in = new_bw_in;
677 new_address->assigned_bw_in.value__ = htonl (new_bw_in);
678 notify |= GNUNET_YES;
680 if (agent->bw_out != new_bw_out)
682 agent->bw_out = new_bw_out;
683 new_address->assigned_bw_out.value__ = htonl (new_bw_out);
684 notify |= GNUNET_YES;
688 if (notify && agent->is_active && (GNUNET_NO == silent))
692 LOG(GNUNET_ERROR_TYPE_DEBUG, " envi_set_active_suggestion() notify\n");
693 agent->suggestion_issue = GNUNET_YES;
694 agent->suggestion_address = new_address;
696 else if (agent->address_inuse)
698 //disconnect case, no new address
699 GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_in.value__));
700 GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_out.value__));
704 agent->suggestion_issue = GNUNET_YES;
705 agent->suggestion_address = agent->address_inuse;
708 agent->address_inuse = new_address;
712 static unsigned long long
713 ril_network_get_assigned (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
715 struct RIL_Peer_Agent *cur;
716 struct RIL_Network *net;
717 unsigned long long sum = 0;
719 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
721 if (cur->is_active && cur->address_inuse)
723 net = cur->address_inuse->solver_information;
724 if (net->type == type)
738 //envi_state_networks (struct GAS_RIL_Handle *solver)
741 // struct RIL_Network net;
742 // int overutilized_in;
743 // int overutilized_out;
745 // for (i = 0; i < solver->networks_count; i++)
747 // net = solver->network_entries[i];
749 // overutilized_in = net.bw_in_assigned > net.bw_in_available;
750 // overutilized_out = net.bw_out_assigned > net.bw_out_available;
752 // solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 0] = ((double) net.bw_in_assigned / (double) net.bw_in_available)*10;
753 // solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 1] = (double) overutilized_in;
754 // solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 2] = ((double) net.bw_out_assigned / (double) net.bw_out_available)*10;
755 // solver->global_state_networks[i * RIL_FEATURES_NETWORK_COUNT + 3] = (double) overutilized_out;
760 * Allocates a state vector and fills it with the features present
761 * @param solver the solver handle
762 * @param agent the agent handle
763 * @return pointer to the state vector
766 envi_get_state (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
770 double *state = GNUNET_malloc (sizeof (double) * agent->m);
771 // struct RIL_Address_Wrapped *cur_address;
772 // const double *preferences;
773 // const double *properties;
774 struct RIL_Network *net;
776 //copy global networks state
777 // for (i = 0; i < solver->networks_count * RIL_FEATURES_NETWORK_COUNT; i++)
779 // state[i] = solver->global_state_networks[i];
782 net = agent->address_inuse->solver_information;
785 state[1] = (double) net->bw_in_assigned / (double) GNUNET_ATS_MaxBandwidth;
786 state[2] = GNUNET_MIN((double) (net->bw_in_available - net->bw_in_assigned), 0) / (double) GNUNET_ATS_MaxBandwidth;
788 // LOG(GNUNET_ERROR_TYPE_INFO, "get_state() state[0] = %f\n", state[0]);
789 // LOG(GNUNET_ERROR_TYPE_INFO, "get_state() state[1] = %f\n", state[1]);
791 // LOG(GNUNET_ERROR_TYPE_INFO, "get_state() W / %08.3f %08.3f \\ \n", agent->W[0][0], agent->W[1][0]);
792 // LOG(GNUNET_ERROR_TYPE_INFO, "get_state() W \\ %08.3f %08.3f / \n", agent->W[0][1], agent->W[1][1]);
796 // preferences = solver->plugin_envi->get_preferences (solver->plugin_envi->get_preference_cls,
798 // for (k = 0; k < GNUNET_ATS_PreferenceCount; k++)
800 // state[i++] = preferences[k];
803 //get address specific features
804 // for (cur_address = agent->addresses_head; NULL != cur_address; cur_address = cur_address->next)
806 // //when changing the number of address specific state features, change RIL_FEATURES_ADDRESS_COUNT macro
807 // state[i++] = cur_address->address_naked->active;
808 // state[i++] = cur_address->address_naked->active ? agent->bw_in : 0;
809 // state[i++] = cur_address->address_naked->active ? agent->bw_out : 0;
810 // properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
811 // cur_address->address_naked);
812 // for (k = 0; k < GNUNET_ATS_QualityPropertiesCount; k++)
814 // state[i++] = properties[k];
822 // * For all networks a peer has an address in, this gets the maximum bandwidth which could
823 // * theoretically be available in one of the networks. This is used for bandwidth normalization.
825 // * @param agent the agent handle
826 // * @param direction_in whether the inbound bandwidth should be considered. Returns the maximum outbound bandwidth if GNUNET_NO
828 //static unsigned long long
829 //ril_get_max_bw (struct RIL_Peer_Agent *agent, int direction_in)
832 // * get the maximum bandwidth possible for a peer, e.g. among all addresses which addresses'
833 // * network could provide the maximum bandwidth if all that bandwidth was used on that one peer.
835 // unsigned long long max = 0;
836 // struct RIL_Address_Wrapped *cur;
837 // struct RIL_Network *net;
839 // for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
841 // net = cur->address_naked->solver_information;
844 // if (net->bw_in_available > max)
846 // max = net->bw_in_available;
851 // if (net->bw_out_available > max)
853 // max = net->bw_out_available;
861 // * Get the index of the quality-property in question
863 // * @param type the quality property type
864 // * @return the index
867 //ril_find_property_index (uint32_t type)
869 // int existing_types[] = GNUNET_ATS_QualityProperties;
871 // for (c = 0; c < GNUNET_ATS_QualityPropertiesCount; c++)
872 // if (existing_types[c] == type)
874 // return GNUNET_SYSERR;
878 //ril_get_atsi (struct ATS_Address *address, uint32_t type)
881 // GNUNET_assert(NULL != address);
883 // if ((NULL == address->atsi) || (0 == address->atsi_count))
886 // for (c1 = 0; c1 < address->atsi_count; c1++)
888 // if (ntohl (address->atsi[c1].type) == type)
889 // return ntohl (address->atsi[c1].value);
895 //envi_reward_global (struct GAS_RIL_Handle *solver)
898 // struct RIL_Network net;
899 // unsigned int sum_in_available = 0;
900 // unsigned int sum_out_available = 0;
901 // unsigned int sum_in_assigned = 0;
902 // unsigned int sum_out_assigned = 0;
906 // for (i = 0; i < solver->networks_count; i++)
908 // net = solver->network_entries[i];
909 // sum_in_available += net.bw_in_available;
910 // sum_in_assigned += net.bw_in_assigned;
911 // sum_out_available += net.bw_out_available;
912 // sum_out_assigned += net.bw_out_assigned;
915 // ratio_in = ((double) sum_in_assigned) / ((double) sum_in_available);
916 // ratio_out = ((double) sum_out_assigned) / ((double) sum_out_available);
918 // // global reward in [1,2]
919 // return ratio_in +1;
920 // return ((ratio_in + ratio_out) / 2) + 1;
924 //envi_reward_local (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
926 // const double *preferences;
927 // const double *properties;
929 // double pref_match = 0;
933 // preferences = solver->plugin_envi->get_preferences (solver->plugin_envi->get_preference_cls,
935 // properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
936 // agent->address_inuse);
939 // prop_index = ril_find_property_index (GNUNET_ATS_QUALITY_NET_DELAY);
940 // dl_norm = 2 - properties[prop_index]; //invert property as we want to maximize for lower latencies
942 // // utilization in [0,1]
943 // bw_norm = (((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_IN)
944 // / (double) ril_get_max_bw (agent, GNUNET_YES))
945 // + ((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_OUT)
946 // / (double) ril_get_max_bw (agent, GNUNET_NO))) / 2;
948 // // preference matching in [0,4]
949 // pref_match += (preferences[GNUNET_ATS_PREFERENCE_LATENCY] * dl_norm);
950 // pref_match += (preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] * bw_norm);
952 // // local reward in [1,2]
953 // return (pref_match / 4) +1;
957 envi_get_collective_utility (struct GAS_RIL_Handle *solver)
959 //TODO! add nash product
960 struct RIL_Peer_Agent *cur;
961 double result = RIL_UTILITY_MAX;
963 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
967 if (cur->address_inuse)
969 result = GNUNET_MIN(result, agent_get_utility(cur));
978 * Gets the reward for the last performed step, which is calculated in equal
979 * parts from the local (the peer specific) and the global (for all peers
982 * @param solver the solver handle
983 * @param agent the agent handle
987 envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
989 struct RIL_Network *net;
991 unsigned long long objective;
993 net = agent->address_inuse->solver_information;
994 if (net->bw_in_assigned > net->bw_in_available)
996 objective = net->bw_in_available - net->bw_in_assigned;
1000 objective = envi_get_collective_utility(solver);
1007 * Doubles the bandwidth for the active address
1009 * @param solver solver handle
1010 * @param agent agent handle
1011 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise the outbound bandwidth
1014 envi_action_bw_double (struct GAS_RIL_Handle *solver,
1015 struct RIL_Peer_Agent *agent,
1018 unsigned long long new_bw;
1022 new_bw = agent->bw_in * 2;
1023 if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
1024 new_bw = GNUNET_ATS_MaxBandwidth;
1025 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1026 agent->bw_out, GNUNET_NO);
1030 new_bw = agent->bw_out * 2;
1031 if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
1032 new_bw = GNUNET_ATS_MaxBandwidth;
1033 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1039 * Cuts the bandwidth for the active address in half. The least amount of bandwidth suggested, is
1040 * the minimum bandwidth for a peer, in order to not invoke a disconnect.
1042 * @param solver solver handle
1043 * @param agent agent handle
1044 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1048 envi_action_bw_halven (struct GAS_RIL_Handle *solver,
1049 struct RIL_Peer_Agent *agent,
1052 unsigned long long new_bw;
1056 new_bw = agent->bw_in / 2;
1057 if (new_bw < MIN_BW || new_bw > agent->bw_in)
1059 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1064 new_bw = agent->bw_out / 2;
1065 if (new_bw < MIN_BW || new_bw > agent->bw_out)
1067 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1073 * Increases the bandwidth by 5 times the minimum bandwidth for the active address.
1075 * @param solver solver handle
1076 * @param agent agent handle
1077 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1081 envi_action_bw_inc (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1083 unsigned long long new_bw;
1087 new_bw = agent->bw_in + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1088 if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
1089 new_bw = GNUNET_ATS_MaxBandwidth;
1090 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1091 agent->bw_out, GNUNET_NO);
1095 new_bw = agent->bw_out + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1096 if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
1097 new_bw = GNUNET_ATS_MaxBandwidth;
1098 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1104 * Decreases the bandwidth by 5 times the minimum bandwidth for the active address. The least amount
1105 * of bandwidth suggested, is the minimum bandwidth for a peer, in order to not invoke a disconnect.
1107 * @param solver solver handle
1108 * @param agent agent handle
1109 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1113 envi_action_bw_dec (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1115 unsigned long long new_bw;
1119 new_bw = agent->bw_in - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1120 if (new_bw < MIN_BW || new_bw > agent->bw_in)
1122 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1127 new_bw = agent->bw_out - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1128 if (new_bw < MIN_BW || new_bw > agent->bw_out)
1130 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1136 * Switches to the address given by its index
1138 * @param solver solver handle
1139 * @param agent agent handle
1140 * @param address_index index of the address as it is saved in the agent's list, starting with zero
1143 envi_action_address_switch (struct GAS_RIL_Handle *solver,
1144 struct RIL_Peer_Agent *agent,
1145 unsigned int address_index)
1147 struct RIL_Address_Wrapped *cur;
1150 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
1152 if (i == address_index)
1154 envi_set_active_suggestion (solver, agent, cur->address_naked, agent->bw_in, agent->bw_out,
1162 //no address with address_index exists, in this case this action should not be callable
1163 GNUNET_assert(GNUNET_NO);
1167 * Puts the action into effect by calling the according function
1169 * @param solver the solver handle
1170 * @param agent the action handle
1171 * @param action the action to perform by the solver
1174 envi_do_action (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int action)
1180 case RIL_ACTION_NOTHING:
1182 case RIL_ACTION_BW_IN_DBL:
1183 envi_action_bw_double (solver, agent, GNUNET_YES);
1185 case RIL_ACTION_BW_IN_HLV:
1186 envi_action_bw_halven (solver, agent, GNUNET_YES);
1188 case RIL_ACTION_BW_IN_INC:
1189 envi_action_bw_inc (solver, agent, GNUNET_YES);
1191 case RIL_ACTION_BW_IN_DEC:
1192 envi_action_bw_dec (solver, agent, GNUNET_YES);
1194 case RIL_ACTION_BW_OUT_DBL:
1195 envi_action_bw_double (solver, agent, GNUNET_NO);
1197 case RIL_ACTION_BW_OUT_HLV:
1198 envi_action_bw_halven (solver, agent, GNUNET_NO);
1200 case RIL_ACTION_BW_OUT_INC:
1201 envi_action_bw_inc (solver, agent, GNUNET_NO);
1203 case RIL_ACTION_BW_OUT_DEC:
1204 envi_action_bw_dec (solver, agent, GNUNET_NO);
1207 if ((action >= RIL_ACTION_TYPE_NUM) && (action < agent->n)) //switch address action
1209 address_index = action - RIL_ACTION_TYPE_NUM;
1211 GNUNET_assert(address_index >= 0);
1213 address_index <= agent_address_get_index (agent, agent->addresses_tail->address_naked));
1215 envi_action_address_switch (solver, agent, address_index);
1218 // error - action does not exist
1219 GNUNET_assert(GNUNET_NO);
1224 * Performs one step of the Markov Decision Process. Other than in the literature the step starts
1225 * after having done the last action a_old. It observes the new state s_next and the reward
1226 * received. Then the coefficient update is done according to the SARSA or Q-learning method. The
1227 * next action is put into effect.
1229 * @param agent the agent performing the step
1232 agent_step (struct RIL_Peer_Agent *agent)
1234 int a_next = RIL_ACTION_INVALID;
1239 LOG(GNUNET_ERROR_TYPE_DEBUG, " agent_step() Peer '%s', algorithm %s\n",
1240 GNUNET_i2s (&agent->peer),
1241 agent->envi->parameters.algorithm ? "Q" : "SARSA");
1243 s_next = envi_get_state (agent->envi, agent);
1244 reward = envi_get_reward (agent->envi, agent);
1245 explore = agent_decide_exploration (agent);
1247 switch (agent->envi->parameters.algorithm)
1249 case RIL_ALGO_SARSA:
1252 a_next = agent_get_action_explore (agent, s_next);
1256 a_next = agent_get_action_best (agent, s_next);
1258 if (RIL_ACTION_INVALID != agent->a_old)
1260 //updates weights with selected action (on-policy), if not first step
1261 agent_update_weights (agent, reward, s_next, a_next);
1262 agent_modify_eligibility (agent, RIL_E_SET, s_next);
1267 a_next = agent_get_action_best (agent, s_next);
1268 if (RIL_ACTION_INVALID != agent->a_old)
1270 //updates weights with best action, disregarding actually selected action (off-policy), if not first step
1271 agent_update_weights (agent, reward, s_next, a_next);
1275 a_next = agent_get_action_explore (agent, s_next);
1276 agent_modify_eligibility (agent, RIL_E_ZERO, NULL);
1280 a_next = agent_get_action_best (agent, s_next);
1281 agent_modify_eligibility (agent, RIL_E_SET, s_next);
1286 GNUNET_assert(RIL_ACTION_INVALID != a_next);
1288 agent_modify_eligibility (agent, RIL_E_ACCUMULATE, s_next);
1290 // GNUNET_log (GNUNET_ERROR_TYPE_INFO, "step() Step# %llu R: %f IN %llu OUT %llu A: %d\n",
1291 // agent->step_count,
1293 // agent->bw_in/1024,
1294 // agent->bw_out/1024,
1297 envi_do_action (agent->envi, agent, a_next);
1299 GNUNET_free(agent->s_old);
1300 agent->s_old = s_next;
1301 agent->a_old = a_next;
1303 agent->step_count += 1;
1307 ril_step (struct GAS_RIL_Handle *solver);
1310 * Task for the scheduler, which performs one step and lets the solver know that
1311 * no further step is scheduled.
1313 * @param cls the solver handle
1314 * @param tc the task context for the scheduler
1317 ril_step_scheduler_task (void *cls, const struct GNUNET_SCHEDULER_TaskContext *tc)
1319 struct GAS_RIL_Handle *solver = cls;
1321 solver->step_next_task_id = GNUNET_SCHEDULER_NO_TASK;
1326 ril_get_used_resource_ratio (struct GAS_RIL_Handle *solver)
1329 struct RIL_Network net;
1330 unsigned long long sum_assigned = 0;
1331 unsigned long long sum_available = 0;
1334 for (i = 0; i < solver->networks_count; i++)
1336 net = solver->network_entries[i];
1337 if (net.bw_in_assigned > 0) //only consider scopes where an address is actually active
1339 sum_assigned += net.bw_in_assigned;
1340 sum_assigned += net.bw_out_assigned;
1341 sum_available += net.bw_in_available;
1342 sum_available += net.bw_out_available;
1345 if (sum_available > 0)
1347 ratio = ((double) sum_assigned) / ((double) sum_available);
1354 return ratio > 1 ? 1 : ratio; //overutilization possible, cap at 1
1358 * Lookup network struct by type
1360 * @param s the solver handle
1361 * @param type the network type
1362 * @return the network struct
1364 static struct RIL_Network *
1365 ril_get_network (struct GAS_RIL_Handle *s, uint32_t type)
1369 for (i = 0; i < s->networks_count; i++)
1371 if (s->network_entries[i].type == type)
1373 return &s->network_entries[i];
1380 ril_network_is_not_full (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
1382 struct RIL_Network *net;
1383 struct RIL_Peer_Agent *agent;
1384 unsigned long long address_count = 0;
1386 for (agent = solver->agents_head; NULL != agent; agent = agent->next)
1388 if (agent->address_inuse && agent->is_active)
1390 net = agent->address_inuse->solver_information;
1391 if (net->type == network)
1398 net = ril_get_network (solver, network);
1399 return (net->bw_in_available > MIN_BW * address_count) && (net->bw_out_available > MIN_BW * address_count);
1403 ril_try_unblock_agent (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int silent)
1405 struct RIL_Address_Wrapped *addr_wrap;
1406 struct RIL_Network *net;
1408 for (addr_wrap = agent->addresses_head; NULL != addr_wrap; addr_wrap = addr_wrap->next)
1410 net = addr_wrap->address_naked->solver_information;
1411 if (ril_network_is_not_full(solver, net->type))
1413 if (NULL == agent->address_inuse)
1414 envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, MIN_BW, MIN_BW, silent);
1418 agent->address_inuse = NULL;
1422 ril_calculate_discount (struct GAS_RIL_Handle *solver)
1424 struct GNUNET_TIME_Absolute time_now;
1425 struct GNUNET_TIME_Relative time_delta;
1428 // MDP case - remove when debugged
1429 if (solver->simulate)
1431 solver->global_discount_variable = solver->parameters.gamma;
1432 solver->global_discount_integrated = 1;
1438 //calculate tau, i.e. how many real valued time units have passed, one time unit is one minimum time step
1439 time_now = GNUNET_TIME_absolute_get ();
1440 time_delta = GNUNET_TIME_absolute_get_difference (solver->step_time_last, time_now);
1441 solver->step_time_last = time_now;
1442 tau = (double) time_delta.rel_value_us
1443 / (double) solver->parameters.step_time_min.rel_value_us;
1445 //calculate reward discounts (once per step for all agents)
1446 solver->global_discount_variable = pow (M_E, ((-1.0) * ((double) solver->parameters.beta) * tau));
1447 solver->global_discount_integrated = (1.0 - solver->global_discount_variable)
1448 / (double) solver->parameters.beta;
1452 ril_calculate_assigned_bwnet (struct GAS_RIL_Handle *solver)
1455 struct RIL_Network *net;
1457 for (c = 0; c < solver->networks_count; c++)
1459 net = &solver->network_entries[c];
1460 net->bw_in_assigned = ril_network_get_assigned(solver, net->type, GNUNET_YES);
1461 net->bw_out_assigned = ril_network_get_assigned(solver, net->type, GNUNET_NO);
1466 * Schedules the next global step in an adaptive way. The more resources are
1467 * left, the earlier the next step is scheduled. This serves the reactivity of
1468 * the solver to changed inputs.
1470 * @param solver the solver handle
1473 ril_step_schedule_next (struct GAS_RIL_Handle *solver)
1479 struct GNUNET_TIME_Relative time_next;
1481 used_ratio = ril_get_used_resource_ratio (solver);
1484 solver->parameters.step_time_min.rel_value_us
1485 <= solver->parameters.step_time_max.rel_value_us);
1487 factor = (double) GNUNET_TIME_relative_subtract (solver->parameters.step_time_max,
1488 solver->parameters.step_time_min).rel_value_us;
1489 offset = (double) solver->parameters.step_time_min.rel_value_us;
1490 y = factor * pow (used_ratio, RIL_INTERVAL_EXPONENT) + offset;
1492 GNUNET_assert(y <= (double ) solver->parameters.step_time_max.rel_value_us);
1493 GNUNET_assert(y >= (double ) solver->parameters.step_time_min.rel_value_us);
1495 time_next = GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MICROSECONDS, (unsigned long long) y);
1497 if (solver->simulate)
1499 time_next = GNUNET_TIME_UNIT_ZERO;
1502 if ((GNUNET_SCHEDULER_NO_TASK == solver->step_next_task_id) && (GNUNET_NO == solver->done))
1504 solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (time_next, &ril_step_scheduler_task,
1510 * Triggers one step per agent
1514 ril_step (struct GAS_RIL_Handle *solver)
1516 struct RIL_Peer_Agent *cur;
1518 if (GNUNET_YES == solver->bulk_lock)
1520 solver->bulk_changes++;
1524 ril_inform (solver, GAS_OP_SOLVE_START, GAS_STAT_SUCCESS);
1526 LOG(GNUNET_ERROR_TYPE_DEBUG, " RIL step number %d\n", solver->step_count);
1528 if (0 == solver->step_count)
1530 solver->step_time_last = GNUNET_TIME_absolute_get ();
1533 ril_calculate_discount (solver);
1534 ril_calculate_assigned_bwnet (solver);
1536 //calculate network state vector
1537 // envi_state_networks(solver);
1539 //trigger one step per active, unblocked agent
1540 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1544 if (NULL == cur->address_inuse)
1546 ril_try_unblock_agent(solver, cur, GNUNET_NO);
1548 if (cur->address_inuse)
1555 ril_calculate_assigned_bwnet (solver);
1557 solver->step_count += 1;
1558 ril_step_schedule_next (solver);
1560 ril_inform (solver, GAS_OP_SOLVE_STOP, GAS_STAT_SUCCESS);
1562 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_START, GAS_STAT_SUCCESS);
1563 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1565 if (cur->suggestion_issue) {
1566 solver->plugin_envi->bandwidth_changed_cb(solver->plugin_envi->bw_changed_cb_cls, cur->suggestion_address);
1567 cur->suggestion_issue = GNUNET_NO;
1570 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_STOP, GAS_STAT_SUCCESS);
1574 ril_count_agents (struct GAS_RIL_Handle *solver)
1577 struct RIL_Peer_Agent *cur_agent;
1579 for (cur_agent = solver->agents_head; NULL != cur_agent; cur_agent = cur_agent->next)
1587 agent_w_start (struct RIL_Peer_Agent *agent)
1590 struct RIL_Peer_Agent *other;
1594 count = ril_count_agents(agent->envi);
1596 for (i = 0; i < agent->n; i++)
1598 for (k = 0; k < agent->m; k++)
1601 agent->W[i][k] = agent->envi->parameters.alpha * (1.0 - 2.0*((double) GNUNET_CRYPTO_random_u32(GNUNET_CRYPTO_QUALITY_WEAK, UINT32_MAX)/(double)UINT32_MAX));
1604 for (other = agent->envi->agents_head; NULL != other; other = other->next)
1606 agent->W[i][k] += (other->W[i][k] / (double) count);
1610 GNUNET_assert(!isinf(agent->W[i][k]));
1616 * Initialize an agent without addresses and its knowledge base
1618 * @param s ril solver
1619 * @param peer the one in question
1620 * @return handle to the new agent
1622 static struct RIL_Peer_Agent *
1623 agent_init (void *s, const struct GNUNET_PeerIdentity *peer)
1626 struct GAS_RIL_Handle * solver = s;
1627 struct RIL_Peer_Agent * agent = GNUNET_malloc (sizeof (struct RIL_Peer_Agent));
1629 agent->envi = solver;
1630 agent->peer = *peer;
1631 agent->step_count = 0;
1632 agent->is_active = GNUNET_NO;
1633 agent->bw_in = MIN_BW;
1634 agent->bw_out = MIN_BW;
1635 agent->suggestion_issue = GNUNET_NO;
1636 agent->n = RIL_ACTION_TYPE_NUM;
1637 agent->m = RIL_FEATURES_INIT_COUNT;
1638 agent->W = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
1639 for (i = 0; i < agent->n; i++)
1641 agent->W[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1643 agent_w_start(agent);
1644 agent->a_old = RIL_ACTION_INVALID;
1645 agent->s_old = GNUNET_malloc (sizeof (double) * agent->m);
1646 agent->e = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1647 agent_modify_eligibility (agent, RIL_E_ZERO, NULL);
1655 * @param solver the solver handle
1656 * @param agent the agent to retire
1659 agent_die (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
1663 for (i = 0; i < agent->n; i++)
1665 GNUNET_free(agent->W[i]);
1667 GNUNET_free(agent->W);
1668 GNUNET_free(agent->e);
1669 GNUNET_free(agent->s_old);
1674 * Returns the agent for a peer
1676 * @param solver the solver handle
1677 * @param peer the identity of the peer
1678 * @param create whether or not to create an agent, if none is allocated yet
1681 static struct RIL_Peer_Agent *
1682 ril_get_agent (struct GAS_RIL_Handle *solver, const struct GNUNET_PeerIdentity *peer, int create)
1684 struct RIL_Peer_Agent *cur;
1686 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1688 if (0 == memcmp (peer, &cur->peer, sizeof(struct GNUNET_PeerIdentity)))
1696 cur = agent_init (solver, peer);
1697 GNUNET_CONTAINER_DLL_insert_tail(solver->agents_head, solver->agents_tail, cur);
1704 * Determine whether at least the minimum bandwidth is set for the network. Otherwise the network is
1705 * considered inactive and not used. Addresses in an inactive network are ignored.
1707 * @param solver solver handle
1708 * @param network the network type
1709 * @return whether or not the network is considered active
1712 ril_network_is_active (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
1714 struct RIL_Network *net;
1716 net = ril_get_network (solver, network);
1717 return net->bw_out_available >= MIN_BW;
1721 * Cuts a slice out of a vector of elements. This is used to decrease the size of the matrix storing
1722 * the reward function approximation. It copies the memory, which is not cut, to the new vector,
1723 * frees the memory of the old vector, and redirects the pointer to the new one.
1725 * @param old pointer to the pointer to the first element of the vector
1726 * @param element_size byte size of the vector elements
1727 * @param hole_start the first element to cut out
1728 * @param hole_length the number of elements to cut out
1729 * @param old_length the length of the old vector
1732 ril_cut_from_vector (void **old,
1733 size_t element_size,
1734 unsigned int hole_start,
1735 unsigned int hole_length,
1736 unsigned int old_length)
1739 char *oldptr = (char *) *old;
1741 unsigned int bytes_before;
1742 unsigned int bytes_hole;
1743 unsigned int bytes_after;
1745 GNUNET_assert(old_length > hole_length);
1746 GNUNET_assert(old_length >= (hole_start + hole_length));
1748 size = element_size * (old_length - hole_length);
1750 bytes_before = element_size * hole_start;
1751 bytes_hole = element_size * hole_length;
1752 bytes_after = element_size * (old_length - hole_start - hole_length);
1760 tmpptr = GNUNET_malloc (size);
1761 memcpy (tmpptr, oldptr, bytes_before);
1762 memcpy (tmpptr + bytes_before, oldptr + (bytes_before + bytes_hole), bytes_after);
1768 *old = (void *) tmpptr;
1772 * Solver API functions
1773 * ---------------------------
1777 * Change relative preference for quality in solver
1779 * @param solver the solver handle
1780 * @param peer the peer to change the preference for
1781 * @param kind the kind to change the preference
1782 * @param pref_rel the normalized preference value for this kind over all clients
1785 GAS_ril_address_change_preference (void *solver,
1786 const struct GNUNET_PeerIdentity *peer,
1787 enum GNUNET_ATS_PreferenceKind kind,
1790 LOG(GNUNET_ERROR_TYPE_DEBUG,
1791 "API_address_change_preference() Preference '%s' for peer '%s' changed to %.2f \n",
1792 GNUNET_ATS_print_preference_type (kind), GNUNET_i2s (peer), pref_rel);
1798 * Entry point for the plugin
1800 * @param cls pointer to the 'struct GNUNET_ATS_PluginEnvironment'
1803 libgnunet_plugin_ats_ril_init (void *cls)
1805 struct GNUNET_ATS_PluginEnvironment *env = cls;
1806 struct GAS_RIL_Handle *solver = GNUNET_new (struct GAS_RIL_Handle);
1807 struct RIL_Network * cur;
1811 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_init() Initializing RIL solver\n");
1813 GNUNET_assert(NULL != env);
1814 GNUNET_assert(NULL != env->cfg);
1815 GNUNET_assert(NULL != env->stats);
1816 GNUNET_assert(NULL != env->bandwidth_changed_cb);
1817 GNUNET_assert(NULL != env->get_preferences);
1818 GNUNET_assert(NULL != env->get_property);
1821 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MIN",
1822 &solver->parameters.step_time_min))
1824 solver->parameters.step_time_min = RIL_DEFAULT_STEP_TIME_MIN;
1827 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MAX",
1828 &solver->parameters.step_time_max))
1830 solver->parameters.step_time_max = RIL_DEFAULT_STEP_TIME_MAX;
1832 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_ALGORITHM", &string))
1834 solver->parameters.algorithm = !strcmp (string, "SARSA") ? RIL_ALGO_SARSA : RIL_ALGO_Q;
1835 GNUNET_free (string);
1839 solver->parameters.algorithm = RIL_DEFAULT_ALGORITHM;
1841 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_BETA", &string))
1843 solver->parameters.beta = strtod (string, NULL);
1844 GNUNET_free (string);
1845 if (!(solver->parameters.beta > 0))
1847 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_DISCOUNT_BETA not configured as positive number. Set to default value of %f instead.\n", RIL_DEFAULT_DISCOUNT_BETA);
1848 solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
1853 solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
1855 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_GAMMA", &string))
1857 solver->parameters.gamma = strtod (string, NULL);
1858 GNUNET_free (string);
1859 if (!(solver->parameters.gamma < 1) || (solver->parameters.gamma < 0))
1861 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_DISCOUNT_GAMMA not configured in range [0,1[. Set to default value of %f instead.\n", RIL_DEFAULT_DISCOUNT_GAMMA);
1862 solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
1867 solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
1870 == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GRADIENT_STEP_SIZE", &string))
1872 solver->parameters.alpha = strtod (string, NULL);
1873 GNUNET_free (string);
1874 if (!(solver->parameters.alpha > 0) || solver->parameters.alpha > 1)
1876 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_GRADIENT_STEP_SIZE not configured in range ]0,1]. Set to default value of %f instead.\n", RIL_DEFAULT_GRADIENT_STEP_SIZE);
1877 solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
1882 solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
1884 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_TRACE_DECAY", &string))
1886 solver->parameters.lambda = strtod (string, NULL);
1887 GNUNET_free (string);
1888 if (solver->parameters.lambda < 0 || solver->parameters.lambda > 1)
1890 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_TRACE_DECAY not configured in range [0,1]. Set to default value of %f instead.\n", RIL_DEFAULT_TRACE_DECAY);
1891 solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
1896 solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
1898 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_EXPLORE_RATIO", &string))
1900 solver->parameters.explore_ratio = strtod (string, NULL);
1901 GNUNET_free (string);
1902 if (solver->parameters.explore_ratio < 0 || solver->parameters.explore_ratio > 1)
1904 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_EXPLORE_RATIO not configured in range [0,1]. Set to default value of %f instead.\n", RIL_DEFAULT_EXPLORE_RATIO);
1905 solver->parameters.explore_ratio = RIL_DEFAULT_EXPLORE_RATIO;
1910 solver->parameters.explore_ratio = RIL_DEFAULT_EXPLORE_RATIO;
1912 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GLOBAL_REWARD_SHARE", &string))
1914 solver->parameters.reward_global_share = strtod (string, NULL);
1915 GNUNET_free (string);
1916 if (solver->parameters.reward_global_share < 0 || solver->parameters.reward_global_share > 1)
1918 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_GLOBAL_REWARD_SHARE not configured in range [0,1]. Set to default value of %f instead.\n", RIL_DEFAULT_GLOBAL_REWARD_SHARE);
1919 solver->parameters.reward_global_share = RIL_DEFAULT_GLOBAL_REWARD_SHARE;
1924 solver->parameters.reward_global_share = RIL_DEFAULT_GLOBAL_REWARD_SHARE;
1926 if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number (env->cfg, "ats", "RIL_SIMULATE", &solver->simulate))
1928 solver->simulate = GNUNET_NO;
1931 env->sf.s_add = &GAS_ril_address_add;
1932 env->sf.s_address_update_property = &GAS_ril_address_property_changed;
1933 env->sf.s_address_update_session = &GAS_ril_address_session_changed;
1934 env->sf.s_address_update_inuse = &GAS_ril_address_inuse_changed;
1935 env->sf.s_address_update_network = &GAS_ril_address_change_network;
1936 env->sf.s_get = &GAS_ril_get_preferred_address;
1937 env->sf.s_get_stop = &GAS_ril_stop_get_preferred_address;
1938 env->sf.s_pref = &GAS_ril_address_change_preference;
1939 env->sf.s_feedback = &GAS_ril_address_preference_feedback;
1940 env->sf.s_del = &GAS_ril_address_delete;
1941 env->sf.s_bulk_start = &GAS_ril_bulk_start;
1942 env->sf.s_bulk_stop = &GAS_ril_bulk_stop;
1944 solver->plugin_envi = env;
1945 solver->networks_count = env->network_count;
1946 solver->network_entries = GNUNET_malloc (env->network_count * sizeof (struct RIL_Network));
1947 solver->step_count = 0;
1948 solver->global_state_networks = GNUNET_malloc (solver->networks_count * RIL_FEATURES_NETWORK_COUNT * sizeof (double));
1949 solver->done = GNUNET_NO;
1951 for (c = 0; c < env->network_count; c++)
1953 cur = &solver->network_entries[c];
1954 cur->type = env->networks[c];
1955 cur->bw_in_available = env->in_quota[c];
1956 cur->bw_out_available = env->out_quota[c];
1957 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);
1960 LOG(GNUNET_ERROR_TYPE_INFO, "init() Parameters:\n");
1961 LOG(GNUNET_ERROR_TYPE_INFO, "init() Algorithm = %s, alpha = %f, beta = %f, lambda = %f\n",
1962 solver->parameters.algorithm ? "Q" : "SARSA",
1963 solver->parameters.alpha,
1964 solver->parameters.beta,
1965 solver->parameters.lambda);
1966 LOG(GNUNET_ERROR_TYPE_INFO, "init() explore = %f, global_share = %f\n",
1967 solver->parameters.explore_ratio,
1968 solver->parameters.reward_global_share);
1974 * Exit point for the plugin
1976 * @param cls the solver handle
1979 libgnunet_plugin_ats_ril_done (void *cls)
1981 struct GAS_RIL_Handle *s = cls;
1982 struct RIL_Peer_Agent *cur_agent;
1983 struct RIL_Peer_Agent *next_agent;
1985 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_done() Shutting down RIL solver\n");
1987 s->done = GNUNET_YES;
1989 cur_agent = s->agents_head;
1990 while (NULL != cur_agent)
1992 next_agent = cur_agent->next;
1993 GNUNET_CONTAINER_DLL_remove(s->agents_head, s->agents_tail, cur_agent);
1994 agent_die (s, cur_agent);
1995 cur_agent = next_agent;
1998 if (GNUNET_SCHEDULER_NO_TASK != s->step_next_task_id)
2000 GNUNET_SCHEDULER_cancel (s->step_next_task_id);
2002 GNUNET_free(s->network_entries);
2003 GNUNET_free(s->global_state_networks);
2010 * Add a new address for a peer to the solver
2012 * The address is already contained in the addresses hashmap!
2014 * @param solver the solver Handle
2015 * @param address the address to add
2016 * @param network network type of this address
2019 GAS_ril_address_add (void *solver, struct ATS_Address *address, uint32_t network)
2021 struct GAS_RIL_Handle *s = solver;
2022 struct RIL_Peer_Agent *agent;
2023 struct RIL_Address_Wrapped *address_wrapped;
2024 struct RIL_Network *net;
2032 LOG (GNUNET_ERROR_TYPE_DEBUG, "API_address_add()\n");
2034 net = ril_get_network (s, network);
2035 address->solver_information = net;
2037 if (!ril_network_is_active (s, network))
2039 LOG(GNUNET_ERROR_TYPE_DEBUG,
2040 "API_address_add() Did not add %s address %s for peer '%s', network does not have enough bandwidth\n",
2041 address->plugin, address->addr, GNUNET_i2s (&address->peer));
2045 agent = ril_get_agent (s, &address->peer, GNUNET_YES);
2048 address_wrapped = GNUNET_malloc (sizeof (struct RIL_Address_Wrapped));
2049 address_wrapped->address_naked = address;
2050 GNUNET_CONTAINER_DLL_insert_tail(agent->addresses_head, agent->addresses_tail, address_wrapped);
2052 //increase size of W
2053 m_new = agent->m + RIL_FEATURES_ADDRESS_COUNT;
2055 n_new = agent->n + 1;
2058 GNUNET_array_grow(agent->W, agent->n, n_new);
2059 for (i = 0; i < n_new; i++)
2064 GNUNET_array_grow(agent->W[i], agent->m, m_new);
2069 GNUNET_array_grow(agent->W[i], zero, m_new);
2073 //increase size of old state vector
2075 GNUNET_array_grow(agent->s_old, agent->m, m_new);
2078 GNUNET_array_grow(agent->e, agent->m, m_new);
2080 ril_try_unblock_agent(s, agent, GNUNET_NO);
2084 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %s for peer '%s'\n",
2085 address->active ? "active" : "inactive", address->plugin, address->addr,
2086 GNUNET_i2s (&address->peer));
2090 * Delete an address in the solver
2092 * The address is not contained in the address hashmap anymore!
2094 * @param solver the solver handle
2095 * @param address the address to remove
2096 * @param session_only delete only session not whole address
2099 GAS_ril_address_delete (void *solver, struct ATS_Address *address, int session_only)
2101 struct GAS_RIL_Handle *s = solver;
2102 struct RIL_Peer_Agent *agent;
2103 struct RIL_Address_Wrapped *address_wrapped;
2104 int address_was_used = address->active;
2109 struct RIL_Network *net;
2111 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_delete() Delete %s%s %s address %s for peer '%s'\n",
2112 session_only ? "session for " : "", address->active ? "active" : "inactive", address->plugin,
2113 address->addr, GNUNET_i2s (&address->peer));
2115 agent = ril_get_agent (s, &address->peer, GNUNET_NO);
2118 net = address->solver_information;
2119 GNUNET_assert(!ril_network_is_active (s, net->type));
2120 LOG(GNUNET_ERROR_TYPE_DEBUG,
2121 "No agent allocated for peer yet, since address was in inactive network\n");
2125 address_index = agent_address_get_index (agent, address);
2126 address_wrapped = agent_address_get (agent, address);
2128 if (NULL == address_wrapped)
2130 net = address->solver_information;
2131 GNUNET_assert(!ril_network_is_active (s, net->type));
2132 LOG(GNUNET_ERROR_TYPE_DEBUG,
2133 "Address not considered by agent, address was in inactive network\n");
2137 GNUNET_CONTAINER_DLL_remove(agent->addresses_head, agent->addresses_tail, address_wrapped);
2138 GNUNET_free(address_wrapped);
2141 m_new = agent->m - RIL_FEATURES_ADDRESS_COUNT;
2142 n_new = agent->n - 1;
2144 LOG(GNUNET_ERROR_TYPE_DEBUG, "first\n");
2146 for (i = 0; i < agent->n; i++)
2148 ril_cut_from_vector ((void **) &agent->W[i], sizeof(double),
2149 //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
2150 ((RIL_FEATURES_INIT_COUNT) //TODO! replace, when adding more networks
2151 + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
2153 GNUNET_free(agent->W[RIL_ACTION_TYPE_NUM + address_index]);
2154 LOG(GNUNET_ERROR_TYPE_DEBUG, "second\n");
2155 ril_cut_from_vector ((void **) &agent->W, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
2157 //correct last action
2158 if (agent->a_old > (RIL_ACTION_TYPE_NUM + address_index))
2162 else if (agent->a_old == (RIL_ACTION_TYPE_NUM + address_index))
2164 agent->a_old = RIL_ACTION_INVALID;
2166 //decrease old state vector and eligibility vector
2167 LOG(GNUNET_ERROR_TYPE_DEBUG, "third\n");
2168 ril_cut_from_vector ((void **) &agent->s_old, sizeof(double),
2169 //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
2170 ((RIL_FEATURES_INIT_COUNT) //TODO! replace when adding more networks
2171 + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
2172 ril_cut_from_vector ((void **) &agent->e, sizeof(double),
2173 //((s->networks_count * RIL_FEATURES_NETWORK_COUNT)
2174 ((RIL_FEATURES_INIT_COUNT) //TODO! replace when adding more networks
2175 + (address_index * RIL_FEATURES_ADDRESS_COUNT)), RIL_FEATURES_ADDRESS_COUNT, agent->m);
2179 if (address_was_used)
2181 if (NULL != agent->addresses_head) //if peer has an address left, use it
2183 envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, MIN_BW, MIN_BW,
2188 envi_set_active_suggestion (s, agent, NULL, 0, 0, GNUNET_NO);
2196 * Update the properties of an address in the solver
2198 * @param solver solver handle
2199 * @param address the address
2200 * @param type the ATSI type in HBO
2201 * @param abs_value the absolute value of the property
2202 * @param rel_value the normalized value
2205 GAS_ril_address_property_changed (void *solver,
2206 struct ATS_Address *address,
2211 LOG(GNUNET_ERROR_TYPE_DEBUG,
2212 "API_address_property_changed() Property '%s' for peer '%s' address %s changed "
2213 "to %.2f \n", GNUNET_ATS_print_property_type (type), GNUNET_i2s (&address->peer),
2214 address->addr, rel_value);
2220 * Update the session of an address in the solver
2222 * NOTE: values in addresses are already updated
2224 * @param solver solver handle
2225 * @param address the address
2226 * @param cur_session the current session
2227 * @param new_session the new session
2230 GAS_ril_address_session_changed (void *solver,
2231 struct ATS_Address *address,
2232 uint32_t cur_session,
2233 uint32_t new_session)
2236 * TODO? Future Work: Potentially add session activity as a feature in state vector
2238 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_session_changed()\n");
2242 * Notify the solver that an address is (not) actively used by transport
2243 * to communicate with a remote peer
2245 * NOTE: values in addresses are already updated
2247 * @param solver solver handle
2248 * @param address the address
2249 * @param in_use usage state
2252 GAS_ril_address_inuse_changed (void *solver, struct ATS_Address *address, int in_use)
2255 * TODO? Future Work: Potentially add usage variable to state vector
2257 LOG(GNUNET_ERROR_TYPE_DEBUG,
2258 "API_address_inuse_changed() Usage for %s address of peer '%s' changed to %s\n",
2259 address->plugin, GNUNET_i2s (&address->peer), (GNUNET_YES == in_use) ? "USED" : "UNUSED");
2263 * Notify solver that the network an address is located in has changed
2265 * NOTE: values in addresses are already updated
2267 * @param solver solver handle
2268 * @param address the address
2269 * @param current_network the current network
2270 * @param new_network the new network
2273 GAS_ril_address_change_network (void *solver,
2274 struct ATS_Address *address,
2275 uint32_t current_network,
2276 uint32_t new_network)
2278 struct GAS_RIL_Handle *s = solver;
2279 struct RIL_Peer_Agent *agent;
2281 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_change_network() Network type changed, moving "
2282 "%s address of peer %s from '%s' to '%s'\n",
2283 (GNUNET_YES == address->active) ? "active" : "inactive", GNUNET_i2s (&address->peer),
2284 GNUNET_ATS_print_network_type (current_network), GNUNET_ATS_print_network_type (new_network));
2286 if (address->active && !ril_network_is_active (solver, new_network))
2288 GAS_ril_address_delete (solver, address, GNUNET_NO);
2292 agent = ril_get_agent (s, &address->peer, GNUNET_NO);
2295 GNUNET_assert(!ril_network_is_active (solver, current_network));
2297 GAS_ril_address_add (s, address, new_network);
2301 address->solver_information = ril_get_network(solver, new_network);
2305 * Give feedback about the current assignment
2307 * @param solver the solver handle
2308 * @param application the application
2309 * @param peer the peer to change the preference for
2310 * @param scope the time interval for this feedback: [now - scope .. now]
2311 * @param kind the kind to change the preference
2312 * @param score the score
2315 GAS_ril_address_preference_feedback (void *solver,
2317 const struct GNUNET_PeerIdentity *peer,
2318 const struct GNUNET_TIME_Relative scope,
2319 enum GNUNET_ATS_PreferenceKind kind,
2322 LOG(GNUNET_ERROR_TYPE_DEBUG,
2323 "API_address_preference_feedback() Peer '%s' got a feedback of %+.3f from application %s for "
2324 "preference %s for %d seconds\n", GNUNET_i2s (peer), "UNKNOWN",
2325 GNUNET_ATS_print_preference_type (kind), scope.rel_value_us / 1000000);
2329 * Start a bulk operation
2331 * @param solver the solver
2334 GAS_ril_bulk_start (void *solver)
2336 struct GAS_RIL_Handle *s = solver;
2338 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_start() lock: %d\n", s->bulk_lock+1);
2344 * Bulk operation done
2346 * @param solver the solver handle
2349 GAS_ril_bulk_stop (void *solver)
2351 struct GAS_RIL_Handle *s = solver;
2353 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_stop() lock: %d\n", s->bulk_lock-1);
2355 if (s->bulk_lock < 1)
2362 if (0 < s->bulk_changes)
2365 s->bulk_changes = 0;
2370 * Tell solver to notify ATS if the address to use changes for a specific
2371 * peer using the bandwidth changed callback
2373 * The solver must only notify about changes for peers with pending address
2376 * @param solver the solver handle
2377 * @param peer the identity of the peer
2379 const struct ATS_Address *
2380 GAS_ril_get_preferred_address (void *solver, const struct GNUNET_PeerIdentity *peer)
2383 * activate agent, return currently chosen address
2385 struct GAS_RIL_Handle *s = solver;
2386 struct RIL_Peer_Agent *agent;
2388 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_get_preferred_address()\n");
2390 agent = ril_get_agent (s, peer, GNUNET_YES);
2392 agent->is_active = GNUNET_YES;
2393 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, agent->bw_out, GNUNET_YES);
2395 ril_try_unblock_agent(solver, agent, GNUNET_YES);
2397 if (agent->address_inuse)
2399 LOG(GNUNET_ERROR_TYPE_DEBUG,
2400 "API_get_preferred_address() Activated agent for peer '%s' with %s address %s\n",
2401 GNUNET_i2s (peer), agent->address_inuse->plugin, agent->address_inuse->addr);
2405 LOG(GNUNET_ERROR_TYPE_DEBUG,
2406 "API_get_preferred_address() Activated agent for peer '%s', but no address available\n",
2410 return agent->address_inuse;
2414 * Tell solver stop notifying ATS about changes for this peers
2416 * The solver must only notify about changes for peers with pending address
2419 * @param solver the solver handle
2420 * @param peer the peer
2423 GAS_ril_stop_get_preferred_address (void *solver, const struct GNUNET_PeerIdentity *peer)
2425 struct GAS_RIL_Handle *s = solver;
2426 struct RIL_Peer_Agent *agent;
2428 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_stop_get_preferred_address()");
2430 agent = ril_get_agent (s, peer, GNUNET_NO);
2437 if (GNUNET_NO == agent->is_active)
2443 agent->is_active = GNUNET_NO;
2445 envi_set_active_suggestion (s, agent, agent->address_inuse, agent->bw_in, agent->bw_out,
2450 LOG(GNUNET_ERROR_TYPE_DEBUG,
2451 "API_stop_get_preferred_address() Paused agent for peer '%s' with %s address\n",
2452 GNUNET_i2s (peer), agent->address_inuse->plugin);
2455 /* end of plugin_ats_ril.c */