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_INTERVAL_EXPONENT 10
35 #define RIL_UTILITY_MAX (double) GNUNET_ATS_MaxBandwidth
37 #define RIL_DEFAULT_STEP_TIME_MIN GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 500)
38 #define RIL_DEFAULT_STEP_TIME_MAX GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 3000)
39 #define RIL_DEFAULT_ALGORITHM RIL_ALGO_SARSA
40 #define RIL_DEFAULT_SELECT RIL_SELECT_EGREEDY
41 #define RIL_DEFAULT_DISCOUNT_BETA 1.0
42 #define RIL_DEFAULT_DISCOUNT_GAMMA 0.5
43 #define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.1
44 #define RIL_DEFAULT_TRACE_DECAY 0.5
45 #define RIL_DEFAULT_EXPLORE_RATIO 0.1
46 #define RIL_DEFAULT_DIVISOR 10
47 #define RIL_DEFAULT_GLOBAL_REWARD_SHARE 0.5
48 #define RIL_DEFAULT_TEMPERATURE 1.0
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 = 2
90 enum RIL_E_Modification
99 * Global learning parameters
101 struct RIL_Learning_Parameters
104 * The TD-algorithm to use
106 enum RIL_Algorithm algorithm;
109 * Gradient-descent step-size
114 * Learning discount variable in the TD-update for semi-MDPs
119 * Learning discount factor in the TD-update for MDPs
124 * Trace-decay factor for eligibility traces
129 * Softmax action-selection temperature
134 * State space divisor
136 unsigned long long int divisor;
139 * Action selection strategy;
141 enum RIL_Select select;
144 * Ratio, with what probability an agent should explore in the e-greed policy
146 double explore_ratio;
149 * How big the share of the global part of the reward signal is
151 double reward_global_share;
154 * Minimal interval time between steps in milliseconds
156 struct GNUNET_TIME_Relative step_time_min;
159 * Maximum interval time between steps in milliseconds
161 struct GNUNET_TIME_Relative step_time_max;
165 * Wrapper for addresses to store them in agent's linked list
167 struct RIL_Address_Wrapped
172 struct RIL_Address_Wrapped *next;
177 struct RIL_Address_Wrapped *prev;
182 struct ATS_Address *address_naked;
185 struct RIL_Peer_Agent
188 * Next agent in solver's linked list
190 struct RIL_Peer_Agent *next;
193 * Previous agent in solver's linked list
195 struct RIL_Peer_Agent *prev;
200 struct GAS_RIL_Handle *envi;
205 struct GNUNET_PeerIdentity peer;
208 * Whether the agent is active or not
213 * Number of performed time-steps
215 unsigned long long step_count;
218 * Experience matrix W
223 * Number of rows of W / Number of state-vector features
228 * Number of columns of W / Number of actions
233 * Last perceived state feature vector
250 struct ATS_Address * address_inuse;
253 * Head of addresses DLL
255 struct RIL_Address_Wrapped * addresses_head;
258 * Tail of addresses DLL
260 struct RIL_Address_Wrapped * addresses_tail;
263 * Inbound bandwidth assigned by the agent
265 unsigned long long bw_in;
268 * Outbound bandwidth assigned by the agent
270 unsigned long long bw_out;
273 * Flag whether a suggestion has to be issued
275 int suggestion_issue;
278 * The address which has to be issued
280 struct ATS_Address * suggestion_address;
288 enum GNUNET_ATS_Network_Type type;
291 * Total available inbound bandwidth
293 unsigned long long bw_in_available;
296 * Bandwidth inbound assigned in network after last step
298 unsigned long long bw_in_assigned;
301 * Total available outbound bandwidth
303 unsigned long long bw_out_available;
306 * * Bandwidth outbound assigned in network after last step
308 unsigned long long bw_out_assigned;
312 * A handle for the reinforcement learning solver
314 struct GAS_RIL_Handle
317 * The solver-plugin environment of the solver-plugin API
319 struct GNUNET_ATS_PluginEnvironment *plugin_envi;
324 struct GNUNET_STATISTICS_Handle *stats;
327 * Number of performed steps
329 unsigned long long step_count;
332 * Timestamp for the last time-step
334 struct GNUNET_TIME_Absolute step_time_last;
337 * Task identifier of the next time-step to be executed
339 GNUNET_SCHEDULER_TaskIdentifier step_next_task_id;
342 * Variable discount factor, dependent on time between steps
344 double global_discount_variable;
347 * Integrated variable discount factor, dependent on time between steps
349 double global_discount_integrated;
352 * Lock for bulk operations
357 * Number of changes during a lock
362 * Learning parameters
364 struct RIL_Learning_Parameters parameters;
367 * Array of networks with global assignment state
369 struct RIL_Network * network_entries;
374 unsigned int networks_count;
377 * List of active peer-agents
379 struct RIL_Peer_Agent * agents_head;
380 struct RIL_Peer_Agent * agents_tail;
388 * Simulate steps, i.e. schedule steps immediately
390 unsigned long long simulate;
395 * ---------------------------
399 ril_count_agents(struct GAS_RIL_Handle * solver);
402 agent_get_utility (struct RIL_Peer_Agent *agent)
404 return (double) agent->bw_in;
408 * Estimate the current action-value for state s and action a
410 * @param agent agent performing the estimation
413 * @return estimation value
416 agent_estimate_q (struct RIL_Peer_Agent *agent, double *state, int action)
421 for (i = 0; i < agent->m; i++)
423 result += state[i] * agent->W[action][i];
426 GNUNET_assert(!isnan(result));
430 return isinf(result) * UINT32_MAX; //TODO! prevent crash when learning diverges
437 * Decide whether to do exploration (i.e. taking a new action) or exploitation (i.e. taking the
438 * currently estimated best action) in the current step
440 * @param agent agent performing the step
441 * @return yes, if exploring
444 agent_decide_exploration (struct RIL_Peer_Agent *agent)
446 //TODO? Future Work: Improve exploration/exploitation trade-off by different mechanisms than e-greedy
447 double r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
448 UINT32_MAX) / (double) UINT32_MAX;
450 if (r < agent->envi->parameters.explore_ratio)
459 * Get the index of the address in the agent's list.
461 * @param agent agent handle
462 * @param address address handle
463 * @return the index, starting with zero
466 agent_address_get_index (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
469 struct RIL_Address_Wrapped *cur;
472 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
475 if (cur->address_naked == address)
483 * Gets the wrapped address from the agent's list
485 * @param agent agent handle
486 * @param address address handle
487 * @return wrapped address
489 static struct RIL_Address_Wrapped *
490 agent_address_get (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
492 struct RIL_Address_Wrapped *cur;
494 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
495 if (cur->address_naked == address)
502 * Gets the action, with the maximal estimated Q-value (i.e. the one currently estimated to bring the
503 * most reward in the future)
505 * @param agent agent performing the calculation
506 * @param state the state from which to take the action
507 * @return the action promising most future reward
510 agent_get_action_best (struct RIL_Peer_Agent *agent, double *state)
513 int max_i = RIL_ACTION_INVALID;
515 double max_q = -DBL_MAX;
517 for (i = 0; i < agent->n; i++)
519 cur_q = agent_estimate_q (agent, state, i);
527 GNUNET_assert(RIL_ACTION_INVALID != max_i);
534 * Gets any action, to explore the action space from that state
536 * @param agent agent performing the calculation
537 * @param state the state from which to take the action
541 agent_get_action_explore (struct RIL_Peer_Agent *agent, double *state)
543 // TODO?: Future Work: Choose the action for exploration, which has been explored the least in this state
544 return GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK, agent->n);
549 * Updates the weights (i.e. coefficients) of the weight vector in matrix W for action a
551 * @param agent the agent performing the update
552 * @param reward the reward received for the last action
553 * @param s_next the new state, the last step got the agent into
554 * @param a_prime the new
557 agent_update_weights (struct RIL_Peer_Agent *agent, double reward, double *s_next, int a_prime)
561 double *theta = agent->W[agent->a_old];
563 delta = agent->envi->global_discount_integrated * reward; //reward
564 delta += agent->envi->global_discount_variable * agent_estimate_q (agent, s_next, a_prime); //discounted future value
565 delta -= agent_estimate_q (agent, agent->s_old, agent->a_old); //one step
567 // 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",
568 // agent->step_count,
569 // agent_estimate_q (agent, agent->s_old, agent->a_old),
570 // agent->envi->parameters.alpha,
572 // agent->envi->global_discount_variable,
573 // agent_estimate_q (agent, s_next, a_prime),
576 for (i = 0; i < agent->m; i++)
578 // LOG(GNUNET_ERROR_TYPE_INFO, "alpha = %f delta = %f e[%d] = %f\n",
579 // agent->envi->parameters.alpha,
583 theta[i] += agent->envi->parameters.alpha * delta * agent->s_old[i];// * agent->E[a_prime][i];
589 * Changes the eligibility trace vector e in various manners:
590 * #RIL_E_ACCUMULATE - adds @a feature to each component as in accumulating eligibility traces
591 * #RIL_E_REPLACE - resets each component to @a feature as in replacing traces
592 * #RIL_E_SET - multiplies e with discount factor and lambda as in the update rule
593 * #RIL_E_ZERO - sets e to 0 as in Watkin's Q-learning algorithm when exploring and when initializing
595 * @param agent the agent handle
596 * @param mod the kind of modification
597 * @param feature the feature vector
600 agent_modify_eligibility (struct RIL_Peer_Agent *agent,
601 enum RIL_E_Modification mod,
608 for (i = 0; i < agent->m; i++)
612 case RIL_E_ACCUMULATE:
613 agent->E[action][i] += feature[i];
616 agent->E[action][i] = agent->E[action][i]+feature[i] > 1 ? 1 : agent->E[action][i]+feature[i]; //TODO? Maybe remove as only accumulating traces really apply
619 agent->E[action][i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
622 for (k = 0; k < agent->n; k++)
633 ril_inform (struct GAS_RIL_Handle *solver,
634 enum GAS_Solver_Operation op,
635 enum GAS_Solver_Status stat)
637 if (NULL != solver->plugin_envi->info_cb)
638 solver->plugin_envi->info_cb (solver->plugin_envi->info_cb_cls, op, stat, GAS_INFO_NONE);
643 * Changes the active assignment suggestion of the handler and invokes the bw_changed callback to
644 * notify ATS of its new decision
646 * @param solver solver handle
647 * @param agent agent handle
648 * @param new_address the address which is to be used
649 * @param new_bw_in the new amount of inbound bandwidth set for this address
650 * @param new_bw_out the new amount of outbound bandwidth set for this address
651 * @param silent disables invocation of the bw_changed callback, if GNUNET_YES
654 envi_set_active_suggestion (struct GAS_RIL_Handle *solver,
655 struct RIL_Peer_Agent *agent,
656 struct ATS_Address *new_address,
657 unsigned long long new_bw_in,
658 unsigned long long new_bw_out,
661 int notify = GNUNET_NO;
663 LOG(GNUNET_ERROR_TYPE_DEBUG, " set_active_suggestion() for peer '%s'\n", GNUNET_i2s (&agent->peer));
666 if (agent->address_inuse != new_address)
668 if (NULL != agent->address_inuse)
670 agent->address_inuse->active = GNUNET_NO;
671 agent->address_inuse->assigned_bw_in.value__ = htonl (0);
672 agent->address_inuse->assigned_bw_out.value__ = htonl (0);
674 if (NULL != new_address)
676 LOG(GNUNET_ERROR_TYPE_DEBUG, " set address active: %s\n", agent->is_active ? "yes" : "no");
677 new_address->active = agent->is_active;
678 new_address->assigned_bw_in.value__ = htonl (agent->bw_in);
679 new_address->assigned_bw_out.value__ = htonl (agent->bw_out);
681 notify |= GNUNET_YES;
687 if (new_address->active != agent->is_active)
689 new_address->active = agent->is_active;
690 notify |= GNUNET_YES;
694 if (agent->bw_in != new_bw_in)
696 agent->bw_in = new_bw_in;
697 new_address->assigned_bw_in.value__ = htonl (new_bw_in);
698 notify |= GNUNET_YES;
700 if (agent->bw_out != new_bw_out)
702 agent->bw_out = new_bw_out;
703 new_address->assigned_bw_out.value__ = htonl (new_bw_out);
704 notify |= GNUNET_YES;
708 if (notify && agent->is_active && (GNUNET_NO == silent))
712 LOG(GNUNET_ERROR_TYPE_DEBUG, " envi_set_active_suggestion() notify\n");
713 agent->suggestion_issue = GNUNET_YES;
714 agent->suggestion_address = new_address;
716 else if (agent->address_inuse)
718 //disconnect case, no new address
719 GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_in.value__));
720 GNUNET_assert(0 == ntohl (agent->address_inuse->assigned_bw_out.value__));
724 agent->suggestion_issue = GNUNET_YES;
725 agent->suggestion_address = agent->address_inuse;
728 agent->address_inuse = new_address;
732 static unsigned long long
733 ril_network_get_assigned (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
735 struct RIL_Peer_Agent *cur;
736 struct RIL_Network *net;
737 unsigned long long sum = 0;
739 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
741 if (cur->is_active && cur->address_inuse)
743 net = cur->address_inuse->solver_information;
744 if (net->type == type)
758 * Allocates a state vector and fills it with the features present
759 * @param solver the solver handle
760 * @param agent the agent handle
761 * @return pointer to the state vector
764 envi_get_state (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
776 state = GNUNET_malloc (sizeof(double) * agent->m);
778 y[0] = (double) agent->bw_out;
779 y[1] = (double) agent->bw_in;
781 m = agent_address_get_index (agent, agent->address_inuse) * (solver->parameters.divisor+1) * (solver->parameters.divisor+1);
782 for (i = 0; i <= solver->parameters.divisor; i++)
784 for (k = 0; k <= solver->parameters.divisor; k++)
786 x[0] = i * GNUNET_ATS_MaxBandwidth / solver->parameters.divisor;
787 x[1] = k * GNUNET_ATS_MaxBandwidth / solver->parameters.divisor;
790 sigma = ((double) GNUNET_ATS_MaxBandwidth / 2) * M_SQRT2;
791 f = exp(-((d[0]*d[0] + d[1]*d[1]) / (2 * sigma * sigma)));
800 // * For all networks a peer has an address in, this gets the maximum bandwidth which could
801 // * theoretically be available in one of the networks. This is used for bandwidth normalization.
803 // * @param agent the agent handle
804 // * @param direction_in whether the inbound bandwidth should be considered. Returns the maximum outbound bandwidth if GNUNET_NO
806 //static unsigned long long
807 //ril_get_max_bw (struct RIL_Peer_Agent *agent, int direction_in)
810 // * get the maximum bandwidth possible for a peer, e.g. among all addresses which addresses'
811 // * network could provide the maximum bandwidth if all that bandwidth was used on that one peer.
813 // unsigned long long max = 0;
814 // struct RIL_Address_Wrapped *cur;
815 // struct RIL_Network *net;
817 // for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
819 // net = cur->address_naked->solver_information;
822 // if (net->bw_in_available > max)
824 // max = net->bw_in_available;
829 // if (net->bw_out_available > max)
831 // max = net->bw_out_available;
839 // * Get the index of the quality-property in question
841 // * @param type the quality property type
842 // * @return the index
845 //ril_find_property_index (uint32_t type)
847 // int existing_types[] = GNUNET_ATS_QualityProperties;
849 // for (c = 0; c < GNUNET_ATS_QualityPropertiesCount; c++)
850 // if (existing_types[c] == type)
852 // return GNUNET_SYSERR;
856 //ril_get_atsi (struct ATS_Address *address, uint32_t type)
859 // GNUNET_assert(NULL != address);
861 // if ((NULL == address->atsi) || (0 == address->atsi_count))
864 // for (c1 = 0; c1 < address->atsi_count; c1++)
866 // if (ntohl (address->atsi[c1].type) == type)
867 // return ntohl (address->atsi[c1].value);
873 //envi_reward_global (struct GAS_RIL_Handle *solver)
876 // struct RIL_Network net;
877 // unsigned int sum_in_available = 0;
878 // unsigned int sum_out_available = 0;
879 // unsigned int sum_in_assigned = 0;
880 // unsigned int sum_out_assigned = 0;
884 // for (i = 0; i < solver->networks_count; i++)
886 // net = solver->network_entries[i];
887 // sum_in_available += net.bw_in_available;
888 // sum_in_assigned += net.bw_in_assigned;
889 // sum_out_available += net.bw_out_available;
890 // sum_out_assigned += net.bw_out_assigned;
893 // ratio_in = ((double) sum_in_assigned) / ((double) sum_in_available);
894 // ratio_out = ((double) sum_out_assigned) / ((double) sum_out_available);
896 // // global reward in [1,2]
897 // return ratio_in +1;
898 // return ((ratio_in + ratio_out) / 2) + 1;
902 //envi_reward_local (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
904 // const double *preferences;
905 // const double *properties;
907 // double pref_match = 0;
911 // preferences = solver->plugin_envi->get_preferences (solver->plugin_envi->get_preference_cls,
913 // properties = solver->plugin_envi->get_property (solver->plugin_envi->get_property_cls,
914 // agent->address_inuse);
917 // prop_index = ril_find_property_index (GNUNET_ATS_QUALITY_NET_DELAY);
918 // dl_norm = 2 - properties[prop_index]; //invert property as we want to maximize for lower latencies
920 // // utilization in [0,1]
921 // bw_norm = (((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_IN)
922 // / (double) ril_get_max_bw (agent, GNUNET_YES))
923 // + ((double) ril_get_atsi (agent->address_inuse, GNUNET_ATS_UTILIZATION_OUT)
924 // / (double) ril_get_max_bw (agent, GNUNET_NO))) / 2;
926 // // preference matching in [0,4]
927 // pref_match += (preferences[GNUNET_ATS_PREFERENCE_LATENCY] * dl_norm);
928 // pref_match += (preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] * bw_norm);
930 // // local reward in [1,2]
931 // return (pref_match / 4) +1;
935 envi_get_collective_utility (struct GAS_RIL_Handle *solver)
937 //TODO! add nash product
938 struct RIL_Peer_Agent *cur;
939 double result = RIL_UTILITY_MAX;
941 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
945 if (cur->address_inuse)
947 result = GNUNET_MIN(result, agent_get_utility(cur));
956 * Gets the reward for the last performed step, which is calculated in equal
957 * parts from the local (the peer specific) and the global (for all peers
960 * @param solver the solver handle
961 * @param agent the agent handle
965 envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
967 struct RIL_Network *net;
969 unsigned long long objective;
971 LOG(GNUNET_ERROR_TYPE_INFO, "address: %x\n", agent->address_inuse);
972 net = agent->address_inuse->solver_information;
973 if (net->bw_in_assigned > net->bw_in_available)
975 objective = net->bw_in_available - net->bw_in_assigned;
979 objective = envi_get_collective_utility(solver);
986 * Doubles the bandwidth for the active address
988 * @param solver solver handle
989 * @param agent agent handle
990 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise the outbound bandwidth
993 envi_action_bw_double (struct GAS_RIL_Handle *solver,
994 struct RIL_Peer_Agent *agent,
997 unsigned long long new_bw;
1001 new_bw = agent->bw_in * 2;
1002 if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
1003 new_bw = GNUNET_ATS_MaxBandwidth;
1004 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1005 agent->bw_out, GNUNET_NO);
1009 new_bw = agent->bw_out * 2;
1010 if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
1011 new_bw = GNUNET_ATS_MaxBandwidth;
1012 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1018 * Cuts the bandwidth for the active address in half. The least amount of bandwidth suggested, is
1019 * the minimum bandwidth for a peer, in order to not invoke a disconnect.
1021 * @param solver solver handle
1022 * @param agent agent handle
1023 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1027 envi_action_bw_halven (struct GAS_RIL_Handle *solver,
1028 struct RIL_Peer_Agent *agent,
1031 unsigned long long new_bw;
1035 new_bw = agent->bw_in / 2;
1036 if (new_bw < MIN_BW || new_bw > agent->bw_in)
1038 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1043 new_bw = agent->bw_out / 2;
1044 if (new_bw < MIN_BW || new_bw > agent->bw_out)
1046 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1052 * Increases the bandwidth by 5 times the minimum bandwidth for the active address.
1054 * @param solver solver handle
1055 * @param agent agent handle
1056 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1060 envi_action_bw_inc (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1062 unsigned long long new_bw;
1066 new_bw = agent->bw_in + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1067 if (new_bw < agent->bw_in || new_bw > GNUNET_ATS_MaxBandwidth)
1068 new_bw = GNUNET_ATS_MaxBandwidth;
1069 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1070 agent->bw_out, GNUNET_NO);
1074 new_bw = agent->bw_out + (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1075 if (new_bw < agent->bw_out || new_bw > GNUNET_ATS_MaxBandwidth)
1076 new_bw = GNUNET_ATS_MaxBandwidth;
1077 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1083 * Decreases the bandwidth by 5 times the minimum bandwidth for the active address. The least amount
1084 * of bandwidth suggested, is the minimum bandwidth for a peer, in order to not invoke a disconnect.
1086 * @param solver solver handle
1087 * @param agent agent handle
1088 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1092 envi_action_bw_dec (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1094 unsigned long long new_bw;
1098 new_bw = agent->bw_in - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1099 if (new_bw < MIN_BW || new_bw > agent->bw_in)
1101 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1106 new_bw = agent->bw_out - (RIL_INC_DEC_STEP_SIZE * MIN_BW);
1107 if (new_bw < MIN_BW || new_bw > agent->bw_out)
1109 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1115 * Switches to the address given by its index
1117 * @param solver solver handle
1118 * @param agent agent handle
1119 * @param address_index index of the address as it is saved in the agent's list, starting with zero
1122 envi_action_address_switch (struct GAS_RIL_Handle *solver,
1123 struct RIL_Peer_Agent *agent,
1124 unsigned int address_index)
1126 struct RIL_Address_Wrapped *cur;
1129 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
1131 if (i == address_index)
1133 envi_set_active_suggestion (solver, agent, cur->address_naked, agent->bw_in, agent->bw_out,
1141 //no address with address_index exists, in this case this action should not be callable
1142 GNUNET_assert(GNUNET_NO);
1146 * Puts the action into effect by calling the according function
1148 * @param solver the solver handle
1149 * @param agent the action handle
1150 * @param action the action to perform by the solver
1153 envi_do_action (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int action)
1159 case RIL_ACTION_NOTHING:
1161 case RIL_ACTION_BW_IN_DBL:
1162 envi_action_bw_double (solver, agent, GNUNET_YES);
1164 case RIL_ACTION_BW_IN_HLV:
1165 envi_action_bw_halven (solver, agent, GNUNET_YES);
1167 case RIL_ACTION_BW_IN_INC:
1168 envi_action_bw_inc (solver, agent, GNUNET_YES);
1170 case RIL_ACTION_BW_IN_DEC:
1171 envi_action_bw_dec (solver, agent, GNUNET_YES);
1173 case RIL_ACTION_BW_OUT_DBL:
1174 envi_action_bw_double (solver, agent, GNUNET_NO);
1176 case RIL_ACTION_BW_OUT_HLV:
1177 envi_action_bw_halven (solver, agent, GNUNET_NO);
1179 case RIL_ACTION_BW_OUT_INC:
1180 envi_action_bw_inc (solver, agent, GNUNET_NO);
1182 case RIL_ACTION_BW_OUT_DEC:
1183 envi_action_bw_dec (solver, agent, GNUNET_NO);
1186 if ((action >= RIL_ACTION_TYPE_NUM) && (action < agent->n)) //switch address action
1188 address_index = action - RIL_ACTION_TYPE_NUM;
1190 GNUNET_assert(address_index >= 0);
1192 address_index <= agent_address_get_index (agent, agent->addresses_tail->address_naked));
1194 envi_action_address_switch (solver, agent, address_index);
1197 // error - action does not exist
1198 GNUNET_assert(GNUNET_NO);
1203 agent_select_egreedy (struct RIL_Peer_Agent *agent, double *state)
1207 if (agent_decide_exploration(agent))
1209 action = agent_get_action_explore(agent, state);
1210 if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
1212 agent_modify_eligibility(agent, RIL_E_ZERO, NULL, action);
1218 action = agent_get_action_best(agent, state);
1219 if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
1221 agent_modify_eligibility(agent, RIL_E_UPDATE, NULL, action);
1228 * Selects the next action with a probability corresponding to its value. The
1229 * probability is calculated using a Boltzmann distribution with a temperature
1230 * value. The higher the temperature, the more are the action selection
1231 * probabilities the same. With a temperature of 0, the selection is greedy,
1232 * i.e. always the action with the highest value is chosen.
1238 agent_select_softmax (struct RIL_Peer_Agent *agent, double *state)
1242 double eqt[agent->n];
1247 a_max = agent_get_action_best(agent, state);
1249 for (i=0; i<agent->n; i++)
1251 eqt[i] = exp(agent_estimate_q(agent,state,i) / agent->envi->parameters.temperature);
1254 for (i=0; i<agent->n; i++)
1258 r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
1259 UINT32_MAX) / (double) UINT32_MAX;
1261 for (i=0; i<agent->n; i++)
1265 if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
1268 agent_modify_eligibility(agent, RIL_E_UPDATE, NULL, i);
1270 agent_modify_eligibility(agent, RIL_E_ZERO, NULL, -1);
1276 GNUNET_assert(GNUNET_NO);
1280 agent_select_action (struct RIL_Peer_Agent *agent, double *state)
1282 if (agent->envi->parameters.select == RIL_SELECT_EGREEDY)
1284 return agent_select_egreedy(agent, state);
1288 return agent_select_softmax(agent, state);
1293 * Performs one step of the Markov Decision Process. Other than in the literature the step starts
1294 * after having done the last action a_old. It observes the new state s_next and the reward
1295 * received. Then the coefficient update is done according to the SARSA or Q-learning method. The
1296 * next action is put into effect.
1298 * @param agent the agent performing the step
1301 agent_step (struct RIL_Peer_Agent *agent)
1303 int a_next = RIL_ACTION_INVALID;
1308 LOG(GNUNET_ERROR_TYPE_DEBUG, " agent_step() Peer '%s', algorithm %s\n",
1309 GNUNET_i2s (&agent->peer),
1310 agent->envi->parameters.algorithm ? "Q" : "SARSA");
1312 s_next = envi_get_state (agent->envi, agent);
1313 reward = envi_get_reward (agent->envi, agent);
1315 switch (agent->envi->parameters.algorithm)
1317 case RIL_ALGO_SARSA:
1318 a_next = agent_select_action (agent, s_next);
1319 if (RIL_ACTION_INVALID != agent->a_old)
1321 //updates weights with selected action (on-policy), if not first step
1322 agent_update_weights (agent, reward, s_next, a_next);
1324 agent_modify_eligibility (agent, RIL_E_UPDATE, s_next, a_next);
1328 a_max = agent_get_action_best (agent, s_next);
1329 if (RIL_ACTION_INVALID != agent->a_old)
1331 //updates weights with best action, disregarding actually selected action (off-policy), if not first step
1332 agent_update_weights (agent, reward, s_next, a_max);
1334 a_next = agent_select_action (agent, s_next);
1338 GNUNET_assert(RIL_ACTION_INVALID != a_next);
1340 agent_modify_eligibility (agent, RIL_E_ACCUMULATE, s_next, a_next);
1342 // GNUNET_log (GNUNET_ERROR_TYPE_INFO, "step() Step# %llu R: %f IN %llu OUT %llu A: %d\n",
1343 // agent->step_count,
1345 // agent->bw_in/1024,
1346 // agent->bw_out/1024,
1349 envi_do_action (agent->envi, agent, a_next);
1351 GNUNET_free(agent->s_old);
1352 agent->s_old = s_next;
1353 agent->a_old = a_next;
1355 agent->step_count += 1;
1359 ril_step (struct GAS_RIL_Handle *solver);
1362 * Task for the scheduler, which performs one step and lets the solver know that
1363 * no further step is scheduled.
1365 * @param cls the solver handle
1366 * @param tc the task context for the scheduler
1369 ril_step_scheduler_task (void *cls, const struct GNUNET_SCHEDULER_TaskContext *tc)
1371 struct GAS_RIL_Handle *solver = cls;
1373 solver->step_next_task_id = GNUNET_SCHEDULER_NO_TASK;
1378 ril_get_used_resource_ratio (struct GAS_RIL_Handle *solver)
1381 struct RIL_Network net;
1382 unsigned long long sum_assigned = 0;
1383 unsigned long long sum_available = 0;
1386 for (i = 0; i < solver->networks_count; i++)
1388 net = solver->network_entries[i];
1389 if (net.bw_in_assigned > 0) //only consider scopes where an address is actually active
1391 sum_assigned += net.bw_in_assigned;
1392 sum_assigned += net.bw_out_assigned;
1393 sum_available += net.bw_in_available;
1394 sum_available += net.bw_out_available;
1397 if (sum_available > 0)
1399 ratio = ((double) sum_assigned) / ((double) sum_available);
1406 return ratio > 1 ? 1 : ratio; //overutilization possible, cap at 1
1410 * Lookup network struct by type
1412 * @param s the solver handle
1413 * @param type the network type
1414 * @return the network struct
1416 static struct RIL_Network *
1417 ril_get_network (struct GAS_RIL_Handle *s, uint32_t type)
1421 for (i = 0; i < s->networks_count; i++)
1423 if (s->network_entries[i].type == type)
1425 return &s->network_entries[i];
1432 ril_network_is_not_full (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
1434 struct RIL_Network *net;
1435 struct RIL_Peer_Agent *agent;
1436 unsigned long long address_count = 0;
1438 for (agent = solver->agents_head; NULL != agent; agent = agent->next)
1440 if (agent->address_inuse && agent->is_active)
1442 net = agent->address_inuse->solver_information;
1443 if (net->type == network)
1450 net = ril_get_network (solver, network);
1451 return (net->bw_in_available > MIN_BW * address_count) && (net->bw_out_available > MIN_BW * address_count);
1455 ril_try_unblock_agent (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int silent)
1457 struct RIL_Address_Wrapped *addr_wrap;
1458 struct RIL_Network *net;
1460 for (addr_wrap = agent->addresses_head; NULL != addr_wrap; addr_wrap = addr_wrap->next)
1462 net = addr_wrap->address_naked->solver_information;
1463 if (ril_network_is_not_full(solver, net->type))
1465 if (NULL == agent->address_inuse)
1466 envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, MIN_BW, MIN_BW, silent);
1470 agent->address_inuse = NULL;
1474 ril_calculate_discount (struct GAS_RIL_Handle *solver)
1476 struct GNUNET_TIME_Absolute time_now;
1477 struct GNUNET_TIME_Relative time_delta;
1480 // MDP case - remove when debugged
1481 if (solver->simulate)
1483 solver->global_discount_variable = solver->parameters.gamma;
1484 solver->global_discount_integrated = 1;
1490 //calculate tau, i.e. how many real valued time units have passed, one time unit is one minimum time step
1491 time_now = GNUNET_TIME_absolute_get ();
1492 time_delta = GNUNET_TIME_absolute_get_difference (solver->step_time_last, time_now);
1493 solver->step_time_last = time_now;
1494 tau = (double) time_delta.rel_value_us
1495 / (double) solver->parameters.step_time_min.rel_value_us;
1497 //calculate reward discounts (once per step for all agents)
1498 solver->global_discount_variable = pow (M_E, ((-1.0) * ((double) solver->parameters.beta) * tau));
1499 solver->global_discount_integrated = (1.0 - solver->global_discount_variable)
1500 / (double) solver->parameters.beta;
1504 ril_calculate_assigned_bwnet (struct GAS_RIL_Handle *solver)
1507 struct RIL_Network *net;
1509 for (c = 0; c < solver->networks_count; c++)
1511 net = &solver->network_entries[c];
1512 net->bw_in_assigned = ril_network_get_assigned(solver, net->type, GNUNET_YES);
1513 net->bw_out_assigned = ril_network_get_assigned(solver, net->type, GNUNET_NO);
1518 * Schedules the next global step in an adaptive way. The more resources are
1519 * left, the earlier the next step is scheduled. This serves the reactivity of
1520 * the solver to changed inputs.
1522 * @param solver the solver handle
1525 ril_step_schedule_next (struct GAS_RIL_Handle *solver)
1531 struct GNUNET_TIME_Relative time_next;
1533 used_ratio = ril_get_used_resource_ratio (solver);
1536 solver->parameters.step_time_min.rel_value_us
1537 <= solver->parameters.step_time_max.rel_value_us);
1539 factor = (double) GNUNET_TIME_relative_subtract (solver->parameters.step_time_max,
1540 solver->parameters.step_time_min).rel_value_us;
1541 offset = (double) solver->parameters.step_time_min.rel_value_us;
1542 y = factor * pow (used_ratio, RIL_INTERVAL_EXPONENT) + offset;
1544 GNUNET_assert(y <= (double ) solver->parameters.step_time_max.rel_value_us);
1545 GNUNET_assert(y >= (double ) solver->parameters.step_time_min.rel_value_us);
1547 time_next = GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MICROSECONDS, (unsigned long long) y);
1549 if (solver->simulate)
1551 time_next = GNUNET_TIME_UNIT_ZERO;
1554 if ((GNUNET_SCHEDULER_NO_TASK == solver->step_next_task_id) && (GNUNET_NO == solver->done))
1556 solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (time_next, &ril_step_scheduler_task,
1562 * Triggers one step per agent
1566 ril_step (struct GAS_RIL_Handle *solver)
1568 struct RIL_Peer_Agent *cur;
1570 if (GNUNET_YES == solver->bulk_lock)
1572 solver->bulk_changes++;
1576 ril_inform (solver, GAS_OP_SOLVE_START, GAS_STAT_SUCCESS);
1578 LOG(GNUNET_ERROR_TYPE_DEBUG, " RIL step number %d\n", solver->step_count);
1580 if (0 == solver->step_count)
1582 solver->step_time_last = GNUNET_TIME_absolute_get ();
1585 ril_calculate_discount (solver);
1586 ril_calculate_assigned_bwnet (solver);
1588 //calculate network state vector
1589 // envi_state_networks(solver);
1591 //trigger one step per active, unblocked agent
1592 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1596 if (NULL == cur->address_inuse)
1598 ril_try_unblock_agent(solver, cur, GNUNET_NO);
1600 if (cur->address_inuse)
1607 ril_calculate_assigned_bwnet (solver);
1609 solver->step_count += 1;
1610 ril_step_schedule_next (solver);
1612 ril_inform (solver, GAS_OP_SOLVE_STOP, GAS_STAT_SUCCESS);
1614 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_START, GAS_STAT_SUCCESS);
1615 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1617 if (cur->suggestion_issue) {
1618 solver->plugin_envi->bandwidth_changed_cb(solver->plugin_envi->bw_changed_cb_cls, cur->suggestion_address);
1619 cur->suggestion_issue = GNUNET_NO;
1622 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_STOP, GAS_STAT_SUCCESS);
1626 ril_count_agents (struct GAS_RIL_Handle *solver)
1629 struct RIL_Peer_Agent *cur_agent;
1631 for (cur_agent = solver->agents_head; NULL != cur_agent; cur_agent = cur_agent->next)
1639 agent_w_start (struct RIL_Peer_Agent *agent)
1642 struct RIL_Peer_Agent *other;
1646 count = ril_count_agents(agent->envi);
1648 for (i = 0; i < agent->n; i++)
1650 for (k = 0; k < agent->m; k++)
1653 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));
1656 for (other = agent->envi->agents_head; NULL != other; other = other->next)
1658 agent->W[i][k] += (other->W[i][k] / (double) count);
1662 GNUNET_assert(!isinf(agent->W[i][k]));
1668 * Initialize an agent without addresses and its knowledge base
1670 * @param s ril solver
1671 * @param peer the one in question
1672 * @return handle to the new agent
1674 static struct RIL_Peer_Agent *
1675 agent_init (void *s, const struct GNUNET_PeerIdentity *peer)
1678 struct GAS_RIL_Handle * solver = s;
1679 struct RIL_Peer_Agent * agent = GNUNET_new (struct RIL_Peer_Agent);
1681 agent->envi = solver;
1682 agent->peer = *peer;
1683 agent->step_count = 0;
1684 agent->is_active = GNUNET_NO;
1685 agent->bw_in = MIN_BW;
1686 agent->bw_out = MIN_BW;
1687 agent->suggestion_issue = GNUNET_NO;
1688 agent->n = RIL_ACTION_TYPE_NUM;
1690 agent->W = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
1691 agent->E = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
1692 for (i = 0; i < agent->n; i++)
1694 agent->W[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1695 agent->E[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1697 agent_w_start(agent);
1698 agent->a_old = RIL_ACTION_INVALID;
1699 agent->s_old = GNUNET_malloc (sizeof (double) * agent->m);
1700 agent->address_inuse = NULL;
1708 * @param solver the solver handle
1709 * @param agent the agent to retire
1712 agent_die (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
1716 for (i = 0; i < agent->n; i++)
1718 GNUNET_free_non_null(agent->W[i]);
1719 GNUNET_free_non_null(agent->E[i]);
1721 GNUNET_free_non_null(agent->W);
1722 GNUNET_free_non_null(agent->E);
1723 GNUNET_free_non_null(agent->s_old);
1728 * Returns the agent for a peer
1730 * @param solver the solver handle
1731 * @param peer the identity of the peer
1732 * @param create whether or not to create an agent, if none is allocated yet
1735 static struct RIL_Peer_Agent *
1736 ril_get_agent (struct GAS_RIL_Handle *solver, const struct GNUNET_PeerIdentity *peer, int create)
1738 struct RIL_Peer_Agent *cur;
1740 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1742 if (0 == memcmp (peer, &cur->peer, sizeof(struct GNUNET_PeerIdentity)))
1750 cur = agent_init (solver, peer);
1751 GNUNET_CONTAINER_DLL_insert_tail(solver->agents_head, solver->agents_tail, cur);
1758 * Determine whether at least the minimum bandwidth is set for the network. Otherwise the network is
1759 * considered inactive and not used. Addresses in an inactive network are ignored.
1761 * @param solver solver handle
1762 * @param network the network type
1763 * @return whether or not the network is considered active
1766 ril_network_is_active (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
1768 struct RIL_Network *net;
1770 net = ril_get_network (solver, network);
1771 return net->bw_out_available >= MIN_BW;
1775 * Cuts a slice out of a vector of elements. This is used to decrease the size of the matrix storing
1776 * the reward function approximation. It copies the memory, which is not cut, to the new vector,
1777 * frees the memory of the old vector, and redirects the pointer to the new one.
1779 * @param old pointer to the pointer to the first element of the vector
1780 * @param element_size byte size of the vector elements
1781 * @param hole_start the first element to cut out
1782 * @param hole_length the number of elements to cut out
1783 * @param old_length the length of the old vector
1786 ril_cut_from_vector (void **old,
1787 size_t element_size,
1788 unsigned int hole_start,
1789 unsigned int hole_length,
1790 unsigned int old_length)
1793 char *oldptr = (char *) *old;
1795 unsigned int bytes_before;
1796 unsigned int bytes_hole;
1797 unsigned int bytes_after;
1799 GNUNET_assert(old_length >= hole_length);
1800 GNUNET_assert(old_length >= (hole_start + hole_length));
1802 size = element_size * (old_length - hole_length);
1804 bytes_before = element_size * hole_start;
1805 bytes_hole = element_size * hole_length;
1806 bytes_after = element_size * (old_length - hole_start - hole_length);
1814 tmpptr = GNUNET_malloc (size);
1815 memcpy (tmpptr, oldptr, bytes_before);
1816 memcpy (tmpptr + bytes_before, oldptr + (bytes_before + bytes_hole), bytes_after);
1822 *old = (void *) tmpptr;
1826 * Solver API functions
1827 * ---------------------------
1831 * Change relative preference for quality in solver
1833 * @param solver the solver handle
1834 * @param peer the peer to change the preference for
1835 * @param kind the kind to change the preference
1836 * @param pref_rel the normalized preference value for this kind over all clients
1839 GAS_ril_address_change_preference (void *solver,
1840 const struct GNUNET_PeerIdentity *peer,
1841 enum GNUNET_ATS_PreferenceKind kind,
1844 LOG(GNUNET_ERROR_TYPE_DEBUG,
1845 "API_address_change_preference() Preference '%s' for peer '%s' changed to %.2f \n",
1846 GNUNET_ATS_print_preference_type (kind), GNUNET_i2s (peer), pref_rel);
1852 * Entry point for the plugin
1854 * @param cls pointer to the 'struct GNUNET_ATS_PluginEnvironment'
1857 libgnunet_plugin_ats_ril_init (void *cls)
1859 struct GNUNET_ATS_PluginEnvironment *env = cls;
1860 struct GAS_RIL_Handle *solver = GNUNET_new (struct GAS_RIL_Handle);
1861 struct RIL_Network * cur;
1865 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_init() Initializing RIL solver\n");
1867 GNUNET_assert(NULL != env);
1868 GNUNET_assert(NULL != env->cfg);
1869 GNUNET_assert(NULL != env->stats);
1870 GNUNET_assert(NULL != env->bandwidth_changed_cb);
1871 GNUNET_assert(NULL != env->get_preferences);
1872 GNUNET_assert(NULL != env->get_property);
1874 if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number(env->cfg, "ats", "RIL_DIVISOR", &solver->parameters.divisor))
1876 solver->parameters.divisor = RIL_DEFAULT_DIVISOR;
1879 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MIN",
1880 &solver->parameters.step_time_min))
1882 solver->parameters.step_time_min = RIL_DEFAULT_STEP_TIME_MIN;
1885 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MAX",
1886 &solver->parameters.step_time_max))
1888 solver->parameters.step_time_max = RIL_DEFAULT_STEP_TIME_MAX;
1890 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_ALGORITHM", &string))
1892 solver->parameters.algorithm = !strcmp (string, "SARSA") ? RIL_ALGO_SARSA : RIL_ALGO_Q;
1893 GNUNET_free (string);
1897 solver->parameters.algorithm = RIL_DEFAULT_ALGORITHM;
1899 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_SELECT", &string))
1901 solver->parameters.select = !strcmp (string, "EGREEDY") ? RIL_SELECT_EGREEDY : RIL_SELECT_SOFTMAX;
1902 GNUNET_free (string);
1906 solver->parameters.select = RIL_DEFAULT_SELECT;
1908 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_BETA", &string))
1910 solver->parameters.beta = strtod (string, NULL);
1911 GNUNET_free (string);
1912 if (!(solver->parameters.beta > 0))
1914 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);
1915 solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
1920 solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
1922 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_DISCOUNT_GAMMA", &string))
1924 solver->parameters.gamma = strtod (string, NULL);
1925 GNUNET_free (string);
1926 if (!(solver->parameters.gamma < 1) || (solver->parameters.gamma < 0))
1928 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);
1929 solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
1934 solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
1937 == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GRADIENT_STEP_SIZE", &string))
1939 solver->parameters.alpha = strtod (string, NULL);
1940 GNUNET_free (string);
1941 if (!(solver->parameters.alpha > 0) || solver->parameters.alpha > 1)
1943 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);
1944 solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
1949 solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
1951 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_TRACE_DECAY", &string))
1953 solver->parameters.lambda = strtod (string, NULL);
1954 GNUNET_free (string);
1955 if (solver->parameters.lambda < 0 || solver->parameters.lambda > 1)
1957 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);
1958 solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
1963 solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
1965 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_EXPLORE_RATIO", &string))
1967 solver->parameters.explore_ratio = strtod (string, NULL);
1968 GNUNET_free (string);
1969 if (solver->parameters.explore_ratio < 0 || solver->parameters.explore_ratio > 1)
1971 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);
1972 solver->parameters.explore_ratio = RIL_DEFAULT_EXPLORE_RATIO;
1977 solver->parameters.explore_ratio = RIL_DEFAULT_EXPLORE_RATIO;
1979 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_GLOBAL_REWARD_SHARE", &string))
1981 solver->parameters.reward_global_share = strtod (string, NULL);
1982 GNUNET_free (string);
1983 if (solver->parameters.reward_global_share < 0 || solver->parameters.reward_global_share > 1)
1985 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);
1986 solver->parameters.reward_global_share = RIL_DEFAULT_GLOBAL_REWARD_SHARE;
1991 solver->parameters.reward_global_share = RIL_DEFAULT_GLOBAL_REWARD_SHARE;
1993 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_TEMPERATURE", &string))
1995 solver->parameters.temperature = strtod (string, NULL);
1996 GNUNET_free (string);
1997 if (solver->parameters.temperature <= 0)
1999 LOG (GNUNET_ERROR_TYPE_WARNING, "RIL_TEMPERATURE not positive. Set to default value of %f instead.\n", RIL_DEFAULT_TEMPERATURE);
2000 solver->parameters.temperature = RIL_DEFAULT_TEMPERATURE;
2005 solver->parameters.temperature = RIL_DEFAULT_TEMPERATURE;
2007 if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number (env->cfg, "ats", "RIL_SIMULATE", &solver->simulate))
2009 solver->simulate = GNUNET_NO;
2012 env->sf.s_add = &GAS_ril_address_add;
2013 env->sf.s_address_update_property = &GAS_ril_address_property_changed;
2014 env->sf.s_address_update_session = &GAS_ril_address_session_changed;
2015 env->sf.s_address_update_inuse = &GAS_ril_address_inuse_changed;
2016 env->sf.s_address_update_network = &GAS_ril_address_change_network;
2017 env->sf.s_get = &GAS_ril_get_preferred_address;
2018 env->sf.s_get_stop = &GAS_ril_stop_get_preferred_address;
2019 env->sf.s_pref = &GAS_ril_address_change_preference;
2020 env->sf.s_feedback = &GAS_ril_address_preference_feedback;
2021 env->sf.s_del = &GAS_ril_address_delete;
2022 env->sf.s_bulk_start = &GAS_ril_bulk_start;
2023 env->sf.s_bulk_stop = &GAS_ril_bulk_stop;
2025 solver->plugin_envi = env;
2026 solver->networks_count = env->network_count;
2027 solver->network_entries = GNUNET_malloc (env->network_count * sizeof (struct RIL_Network));
2028 solver->step_count = 0;
2029 solver->done = GNUNET_NO;
2031 for (c = 0; c < env->network_count; c++)
2033 cur = &solver->network_entries[c];
2034 cur->type = env->networks[c];
2035 cur->bw_in_available = env->in_quota[c];
2036 cur->bw_out_available = env->out_quota[c];
2037 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);
2040 LOG(GNUNET_ERROR_TYPE_INFO, "init() Parameters:\n");
2041 LOG(GNUNET_ERROR_TYPE_INFO, "init() Algorithm = %s, alpha = %f, beta = %f, lambda = %f\n",
2042 solver->parameters.algorithm ? "Q" : "SARSA",
2043 solver->parameters.alpha,
2044 solver->parameters.beta,
2045 solver->parameters.lambda);
2046 LOG(GNUNET_ERROR_TYPE_INFO, "init() explore = %f, global_share = %f\n",
2047 solver->parameters.explore_ratio,
2048 solver->parameters.reward_global_share);
2054 * Exit point for the plugin
2056 * @param cls the solver handle
2059 libgnunet_plugin_ats_ril_done (void *cls)
2061 struct GAS_RIL_Handle *s = cls;
2062 struct RIL_Peer_Agent *cur_agent;
2063 struct RIL_Peer_Agent *next_agent;
2065 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_done() Shutting down RIL solver\n");
2067 s->done = GNUNET_YES;
2069 cur_agent = s->agents_head;
2070 while (NULL != cur_agent)
2072 next_agent = cur_agent->next;
2073 GNUNET_CONTAINER_DLL_remove(s->agents_head, s->agents_tail, cur_agent);
2074 agent_die (s, cur_agent);
2075 cur_agent = next_agent;
2078 if (GNUNET_SCHEDULER_NO_TASK != s->step_next_task_id)
2080 GNUNET_SCHEDULER_cancel (s->step_next_task_id);
2082 GNUNET_free(s->network_entries);
2089 * Add a new address for a peer to the solver
2091 * The address is already contained in the addresses hashmap!
2093 * @param solver the solver Handle
2094 * @param address the address to add
2095 * @param network network type of this address
2098 GAS_ril_address_add (void *solver, struct ATS_Address *address, uint32_t network)
2100 struct GAS_RIL_Handle *s = solver;
2101 struct RIL_Peer_Agent *agent;
2102 struct RIL_Address_Wrapped *address_wrapped;
2103 struct RIL_Network *net;
2111 LOG (GNUNET_ERROR_TYPE_DEBUG, "API_address_add()\n");
2113 net = ril_get_network (s, network);
2114 address->solver_information = net;
2116 if (!ril_network_is_active (s, network))
2118 LOG(GNUNET_ERROR_TYPE_DEBUG,
2119 "API_address_add() Did not add %s address %s for peer '%s', network does not have enough bandwidth\n",
2120 address->plugin, address->addr, GNUNET_i2s (&address->peer));
2124 agent = ril_get_agent (s, &address->peer, GNUNET_YES);
2127 address_wrapped = GNUNET_new (struct RIL_Address_Wrapped);
2128 address_wrapped->address_naked = address;
2129 GNUNET_CONTAINER_DLL_insert_tail(agent->addresses_head, agent->addresses_tail, address_wrapped);
2131 //increase size of W
2132 m_new = agent->m + ((s->parameters.divisor+1) * (s->parameters.divisor+1));
2134 n_new = agent->n + 1;
2137 GNUNET_array_grow(agent->W, agent->n, n_new);
2139 GNUNET_array_grow(agent->E, agent->n, n_new);
2140 for (i = 0; i < n_new; i++)
2145 GNUNET_array_grow(agent->W[i], agent->m, m_new);
2147 GNUNET_array_grow(agent->E[i], agent->m, m_new);
2152 GNUNET_array_grow(agent->W[i], zero, m_new);
2154 GNUNET_array_grow(agent->E[i], zero, m_new);
2158 //increase size of old state vector
2160 GNUNET_array_grow(agent->s_old, agent->m, m_new);
2162 ril_try_unblock_agent(s, agent, GNUNET_NO);
2166 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %s for peer '%s'\n",
2167 address->active ? "active" : "inactive", address->plugin, address->addr,
2168 GNUNET_i2s (&address->peer));
2172 * Delete an address in the solver
2174 * The address is not contained in the address hashmap anymore!
2176 * @param solver the solver handle
2177 * @param address the address to remove
2178 * @param session_only delete only session not whole address
2181 GAS_ril_address_delete (void *solver, struct ATS_Address *address, int session_only)
2183 struct GAS_RIL_Handle *s = solver;
2184 struct RIL_Peer_Agent *agent;
2185 struct RIL_Address_Wrapped *address_wrapped;
2186 int address_was_used = address->active;
2191 struct RIL_Network *net;
2193 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_delete() Delete %s%s %s address %s for peer '%s'\n",
2194 session_only ? "session for " : "", address->active ? "active" : "inactive", address->plugin,
2195 address->addr, GNUNET_i2s (&address->peer));
2197 agent = ril_get_agent (s, &address->peer, GNUNET_NO);
2200 net = address->solver_information;
2201 GNUNET_assert(!ril_network_is_active (s, net->type));
2202 LOG(GNUNET_ERROR_TYPE_DEBUG,
2203 "No agent allocated for peer yet, since address was in inactive network\n");
2207 address_index = agent_address_get_index (agent, address);
2208 address_wrapped = agent_address_get (agent, address);
2210 if (NULL == address_wrapped)
2212 net = address->solver_information;
2213 GNUNET_assert(!ril_network_is_active (s, net->type));
2214 LOG(GNUNET_ERROR_TYPE_DEBUG,
2215 "Address not considered by agent, address was in inactive network\n");
2219 GNUNET_CONTAINER_DLL_remove(agent->addresses_head, agent->addresses_tail, address_wrapped);
2220 GNUNET_free(address_wrapped);
2223 m_new = agent->m - ((s->parameters.divisor+1) * (s->parameters.divisor+1));
2224 n_new = agent->n - 1;
2226 for (i = 0; i < agent->n; i++)
2228 LOG(GNUNET_ERROR_TYPE_DEBUG, "first\n");
2229 ril_cut_from_vector ((void **) &agent->W[i], sizeof(double),
2230 address_index * ((s->parameters.divisor+1) * (s->parameters.divisor+1)),
2231 ((s->parameters.divisor+1) * (s->parameters.divisor+1)), agent->m);
2232 LOG(GNUNET_ERROR_TYPE_DEBUG, "sec\n");
2233 ril_cut_from_vector ((void **) &agent->E[i], sizeof(double),
2234 address_index * ((s->parameters.divisor+1) * (s->parameters.divisor+1)),
2235 ((s->parameters.divisor+1) * (s->parameters.divisor+1)), agent->m);
2237 GNUNET_free_non_null(agent->W[RIL_ACTION_TYPE_NUM + address_index]);
2238 GNUNET_free_non_null(agent->E[RIL_ACTION_TYPE_NUM + address_index]);
2239 LOG(GNUNET_ERROR_TYPE_DEBUG, "third\n");
2240 ril_cut_from_vector ((void **) &agent->W, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
2242 LOG(GNUNET_ERROR_TYPE_DEBUG, "fourth\n");
2243 ril_cut_from_vector ((void **) &agent->E, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
2245 //correct last action
2246 if (agent->a_old > (RIL_ACTION_TYPE_NUM + address_index))
2250 else if (agent->a_old == (RIL_ACTION_TYPE_NUM + address_index))
2252 agent->a_old = RIL_ACTION_INVALID;
2254 //decrease old state vector
2255 LOG(GNUNET_ERROR_TYPE_DEBUG, "fifth\n");
2256 ril_cut_from_vector ((void **) &agent->s_old, sizeof(double),
2257 address_index * ((s->parameters.divisor+1) * (s->parameters.divisor+1)),
2258 ((s->parameters.divisor+1) * (s->parameters.divisor+1)), agent->m);
2262 if (address_was_used)
2264 if (NULL != agent->addresses_head) //if peer has an address left, use it
2266 envi_set_active_suggestion (s, agent, agent->addresses_head->address_naked, MIN_BW, MIN_BW,
2271 envi_set_active_suggestion (s, agent, NULL, 0, 0, GNUNET_NO);
2279 * Update the properties of an address in the solver
2281 * @param solver solver handle
2282 * @param address the address
2283 * @param type the ATSI type in HBO
2284 * @param abs_value the absolute value of the property
2285 * @param rel_value the normalized value
2288 GAS_ril_address_property_changed (void *solver,
2289 struct ATS_Address *address,
2294 LOG(GNUNET_ERROR_TYPE_DEBUG,
2295 "API_address_property_changed() Property '%s' for peer '%s' address %s changed "
2296 "to %.2f \n", GNUNET_ATS_print_property_type (type), GNUNET_i2s (&address->peer),
2297 address->addr, rel_value);
2303 * Update the session of an address in the solver
2305 * NOTE: values in addresses are already updated
2307 * @param solver solver handle
2308 * @param address the address
2309 * @param cur_session the current session
2310 * @param new_session the new session
2313 GAS_ril_address_session_changed (void *solver,
2314 struct ATS_Address *address,
2315 uint32_t cur_session,
2316 uint32_t new_session)
2319 * TODO? Future Work: Potentially add session activity as a feature in state vector
2321 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_session_changed()\n");
2325 * Notify the solver that an address is (not) actively used by transport
2326 * to communicate with a remote peer
2328 * NOTE: values in addresses are already updated
2330 * @param solver solver handle
2331 * @param address the address
2332 * @param in_use usage state
2335 GAS_ril_address_inuse_changed (void *solver, struct ATS_Address *address, int in_use)
2338 * TODO? Future Work: Potentially add usage variable to state vector
2340 LOG(GNUNET_ERROR_TYPE_DEBUG,
2341 "API_address_inuse_changed() Usage for %s address of peer '%s' changed to %s\n",
2342 address->plugin, GNUNET_i2s (&address->peer), (GNUNET_YES == in_use) ? "USED" : "UNUSED");
2346 * Notify solver that the network an address is located in has changed
2348 * NOTE: values in addresses are already updated
2350 * @param solver solver handle
2351 * @param address the address
2352 * @param current_network the current network
2353 * @param new_network the new network
2356 GAS_ril_address_change_network (void *solver,
2357 struct ATS_Address *address,
2358 uint32_t current_network,
2359 uint32_t new_network)
2361 struct GAS_RIL_Handle *s = solver;
2362 struct RIL_Peer_Agent *agent;
2364 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_change_network() Network type changed, moving "
2365 "%s address of peer %s from '%s' to '%s'\n",
2366 (GNUNET_YES == address->active) ? "active" : "inactive", GNUNET_i2s (&address->peer),
2367 GNUNET_ATS_print_network_type (current_network), GNUNET_ATS_print_network_type (new_network));
2369 if (address->active && !ril_network_is_active (solver, new_network))
2371 GAS_ril_address_delete (solver, address, GNUNET_NO);
2375 agent = ril_get_agent (s, &address->peer, GNUNET_NO);
2378 GNUNET_assert(!ril_network_is_active (solver, current_network));
2380 GAS_ril_address_add (s, address, new_network);
2384 address->solver_information = ril_get_network(solver, new_network);
2388 * Give feedback about the current assignment
2390 * @param solver the solver handle
2391 * @param application the application
2392 * @param peer the peer to change the preference for
2393 * @param scope the time interval for this feedback: [now - scope .. now]
2394 * @param kind the kind to change the preference
2395 * @param score the score
2398 GAS_ril_address_preference_feedback (void *solver,
2400 const struct GNUNET_PeerIdentity *peer,
2401 const struct GNUNET_TIME_Relative scope,
2402 enum GNUNET_ATS_PreferenceKind kind,
2405 LOG(GNUNET_ERROR_TYPE_DEBUG,
2406 "API_address_preference_feedback() Peer '%s' got a feedback of %+.3f from application %s for "
2407 "preference %s for %d seconds\n", GNUNET_i2s (peer), "UNKNOWN",
2408 GNUNET_ATS_print_preference_type (kind), scope.rel_value_us / 1000000);
2412 * Start a bulk operation
2414 * @param solver the solver
2417 GAS_ril_bulk_start (void *solver)
2419 struct GAS_RIL_Handle *s = solver;
2421 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_start() lock: %d\n", s->bulk_lock+1);
2427 * Bulk operation done
2429 * @param solver the solver handle
2432 GAS_ril_bulk_stop (void *solver)
2434 struct GAS_RIL_Handle *s = solver;
2436 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_bulk_stop() lock: %d\n", s->bulk_lock-1);
2438 if (s->bulk_lock < 1)
2445 if (0 < s->bulk_changes)
2448 s->bulk_changes = 0;
2453 * Tell solver to notify ATS if the address to use changes for a specific
2454 * peer using the bandwidth changed callback
2456 * The solver must only notify about changes for peers with pending address
2459 * @param solver the solver handle
2460 * @param peer the identity of the peer
2462 const struct ATS_Address *
2463 GAS_ril_get_preferred_address (void *solver, const struct GNUNET_PeerIdentity *peer)
2466 * activate agent, return currently chosen address
2468 struct GAS_RIL_Handle *s = solver;
2469 struct RIL_Peer_Agent *agent;
2471 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_get_preferred_address()\n");
2473 agent = ril_get_agent (s, peer, GNUNET_YES);
2475 agent->is_active = GNUNET_YES;
2476 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, agent->bw_out, GNUNET_YES);
2478 ril_try_unblock_agent(solver, agent, GNUNET_YES);
2480 if (agent->address_inuse)
2482 LOG(GNUNET_ERROR_TYPE_DEBUG,
2483 "API_get_preferred_address() Activated agent for peer '%s' with %s address %s\n",
2484 GNUNET_i2s (peer), agent->address_inuse->plugin, agent->address_inuse->addr);
2488 LOG(GNUNET_ERROR_TYPE_DEBUG,
2489 "API_get_preferred_address() Activated agent for peer '%s', but no address available\n",
2493 return agent->address_inuse;
2497 * Tell solver stop notifying ATS about changes for this peers
2499 * The solver must only notify about changes for peers with pending address
2502 * @param solver the solver handle
2503 * @param peer the peer
2506 GAS_ril_stop_get_preferred_address (void *solver, const struct GNUNET_PeerIdentity *peer)
2508 struct GAS_RIL_Handle *s = solver;
2509 struct RIL_Peer_Agent *agent;
2511 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_stop_get_preferred_address()");
2513 agent = ril_get_agent (s, peer, GNUNET_NO);
2520 if (GNUNET_NO == agent->is_active)
2526 agent->is_active = GNUNET_NO;
2528 envi_set_active_suggestion (s, agent, agent->address_inuse, agent->bw_in, agent->bw_out,
2533 LOG(GNUNET_ERROR_TYPE_DEBUG,
2534 "API_stop_get_preferred_address() Paused agent for peer '%s' with %s address\n",
2535 GNUNET_i2s (peer), agent->address_inuse->plugin);
2538 /* end of plugin_ats_ril.c */