2 This file is part of GNUnet.
3 Copyright (C) 2011-2014 GNUnet e.V.
5 GNUnet is free software: you can redistribute it and/or modify it
6 under the terms of the GNU Affero General Public License as published
7 by the Free Software Foundation, either version 3 of the License,
8 or (at your 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 Affero General Public License for more details.
15 You should have received a copy of the GNU Affero General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
20 * @file ats/plugin_ats_ril.c
21 * @brief ATS reinforcement learning solver
22 * @author Fabian Oehlmann
23 * @author Matthias Wachs
28 #include "gnunet_ats_plugin.h"
29 #include "gnunet-service-ats_addresses.h"
33 #define LOG(kind,...) GNUNET_log_from (kind, "ats-ril",__VA_ARGS__)
35 #define RIL_MIN_BW (5 * ntohl (GNUNET_CONSTANTS_DEFAULT_BW_IN_OUT.value__))
36 #define RIL_MAX_BW GNUNET_ATS_MaxBandwidth
38 #define RIL_ACTION_INVALID -1
39 #define RIL_INTERVAL_EXPONENT 10
40 #define RIL_UTILITY_DELAY_MAX 1000
42 #define RIL_DEFAULT_STEP_TIME_MIN GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 200)
43 #define RIL_DEFAULT_STEP_TIME_MAX GNUNET_TIME_relative_multiply (GNUNET_TIME_UNIT_MILLISECONDS, 2000)
44 #define RIL_DEFAULT_ALGORITHM RIL_ALGO_SARSA
45 #define RIL_DEFAULT_SELECT RIL_SELECT_SOFTMAX
46 #define RIL_DEFAULT_WELFARE RIL_WELFARE_NASH
47 #define RIL_DEFAULT_DISCOUNT_BETA 0.6
48 #define RIL_DEFAULT_DISCOUNT_GAMMA 0.5
49 #define RIL_DEFAULT_GRADIENT_STEP_SIZE 0.01
50 #define RIL_DEFAULT_TRACE_DECAY 0.5
51 #define RIL_DEFAULT_EXPLORE_RATIO 1
52 #define RIL_DEFAULT_EXPLORE_DECAY 0.95
53 #define RIL_DEFAULT_RBF_DIVISOR 50
54 #define RIL_DEFAULT_TEMPERATURE 0.1
55 #define RIL_DEFAULT_TEMPERATURE_DECAY 1
57 #define RIL_INC_DEC_STEP_SIZE 1
58 #define RIL_NOP_DECAY 0.5
61 * ATS reinforcement learning solver
67 * The actions, how an agent can manipulate the current assignment. I.e. how the bandwidth can be
68 * changed for the currently chosen address. Not depicted in the enum are the actions of switching
69 * to a particular address. The action of switching to address with index i is depicted by the
70 * number (RIL_ACTION_TYPE_NUM + i).
74 RIL_ACTION_NOTHING = 0,
75 RIL_ACTION_BW_IN_DBL = -2, //TODO? Potentially add more actions
76 RIL_ACTION_BW_IN_HLV = -3,
77 RIL_ACTION_BW_IN_INC = 1,
78 RIL_ACTION_BW_IN_DEC = 2,
79 RIL_ACTION_BW_OUT_DBL = -4,
80 RIL_ACTION_BW_OUT_HLV = -5,
81 RIL_ACTION_BW_OUT_INC = 3,
82 RIL_ACTION_BW_OUT_DEC = 4,
83 RIL_ACTION_TYPE_NUM = 5
94 RIL_SELECT_SOFTMAX = 0,
95 RIL_SELECT_EGREEDY = 1
101 RIL_WELFARE_EGALITARIAN
104 enum RIL_E_Modification
113 * Global learning parameters
115 struct RIL_Learning_Parameters
118 * The TD-algorithm to use
120 enum RIL_Algorithm algorithm;
123 * Gradient-descent step-size
128 * Learning discount variable in the TD-update for semi-MDPs
133 * Learning discount factor in the TD-update for MDPs
138 * Trace-decay factor for eligibility traces
143 * Whether to accumulate or replace eligibility traces
145 enum RIL_E_Modification eligibility_trace_mode;
148 * Initial softmax action-selection temperature
150 double temperature_init;
153 * Softmax action-selection temperature
158 * Decay factor of the temperature value
160 double temperature_decay;
163 * Which measure of social welfare should be used
165 enum RIL_Welfare social_welfare;
168 * State space divisor
170 unsigned long long rbf_divisor;
173 * Action selection strategy;
175 enum RIL_Select select;
178 * Initial exploration ratio value
183 * Ratio, with what probability an agent should explore in the e-greed policy
188 * Decay factor of the explore ratio
190 double epsilon_decay;
193 * Minimal interval time between steps in milliseconds
195 struct GNUNET_TIME_Relative step_time_min;
198 * Maximum interval time between steps in milliseconds
200 struct GNUNET_TIME_Relative step_time_max;
204 * Wrapper for addresses to store them in agent's linked list
206 struct RIL_Address_Wrapped
211 struct RIL_Address_Wrapped *next;
216 struct RIL_Address_Wrapped *prev;
221 struct ATS_Address *address_naked;
225 struct RIL_Peer_Agent
228 * Next agent in solver's linked list
230 struct RIL_Peer_Agent *next;
233 * Previous agent in solver's linked list
235 struct RIL_Peer_Agent *prev;
240 struct GAS_RIL_Handle *envi;
245 struct GNUNET_PeerIdentity peer;
248 * Whether the agent is active or not
253 * Number of performed time-steps
255 unsigned long long step_count;
258 * Experience matrix W
263 * Number of rows of W / Number of state-vector features
268 * Number of columns of W / Number of actions
273 * Last perceived state feature vector
288 * Whether to reset the eligibility traces to 0 after a Q-exploration step
290 int eligibility_reset;
295 struct ATS_Address *address_inuse;
298 * Head of addresses DLL
300 struct RIL_Address_Wrapped *addresses_head;
303 * Tail of addresses DLL
305 struct RIL_Address_Wrapped *addresses_tail;
308 * Inbound bandwidth assigned by the agent
313 * Outbound bandwidth assigned by the agent
318 * Flag whether a suggestion has to be issued
320 int suggestion_issue;
323 * The address which has to be issued
325 struct ATS_Address *suggestion_address;
328 * The agent's last objective value
330 double objective_old;
343 enum GNUNET_ATS_Network_Type type;
346 * Total available inbound bandwidth
348 uint32_t bw_in_available;
351 * Bandwidth inbound assigned in network after last step
353 uint32_t bw_in_assigned;
356 * Bandwidth inbound actually utilized in the network
358 uint32_t bw_in_utilized;
361 * Total available outbound bandwidth
363 uint32_t bw_out_available;
366 * Bandwidth outbound assigned in network after last step
368 unsigned long long bw_out_assigned;
371 * Bandwidth outbound actually utilized in the network
373 unsigned long long bw_out_utilized;
376 * Number of active agents in scope
378 unsigned int active_agent_count;
381 * The social welfare achieved in the scope
383 double social_welfare;
387 * A handle for the reinforcement learning solver
389 struct GAS_RIL_Handle
392 * The solver-plugin environment of the solver-plugin API
394 struct GNUNET_ATS_PluginEnvironment *env;
397 * Number of performed steps
399 unsigned long long step_count;
402 * Timestamp for the last time-step
404 struct GNUNET_TIME_Absolute step_time_last;
407 * Task identifier of the next time-step to be executed
409 struct GNUNET_SCHEDULER_Task * step_next_task_id;
412 * Variable discount factor, dependent on time between steps
414 double global_discount_variable;
417 * Integrated variable discount factor, dependent on time between steps
419 double global_discount_integrated;
422 * Lock for bulk operations
427 * Number of changes during a lock
432 * Learning parameters
434 struct RIL_Learning_Parameters parameters;
437 * Array of networks with global assignment state
439 struct RIL_Scope * network_entries;
444 unsigned int networks_count;
447 * List of active peer-agents
449 struct RIL_Peer_Agent * agents_head;
450 struct RIL_Peer_Agent * agents_tail;
458 * Simulate steps, i.e. schedule steps immediately
460 unsigned long long simulate;
464 * "Private" functions
465 * ---------------------------
469 * Estimate the current action-value for state s and action a
471 * @param agent agent performing the estimation
474 * @return estimation value
477 agent_q (struct RIL_Peer_Agent *agent,
484 for (i = 0; i < agent->m; i++)
485 result += state[i] * agent->W[action][i];
487 /* prevent crashes if learning diverges */
489 return isnan(result) * UINT32_MAX;
491 return isinf(result) * UINT32_MAX;
497 * Get the index of the address in the agent's list.
499 * @param agent agent handle
500 * @param address address handle
501 * @return the index, starting with zero
504 agent_address_get_index (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
507 struct RIL_Address_Wrapped *cur;
510 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
513 if (cur->address_naked == address)
521 * Gets the wrapped address from the agent's list
523 * @param agent agent handle
524 * @param address address handle
525 * @return wrapped address
527 static struct RIL_Address_Wrapped *
528 agent_address_get_wrapped (struct RIL_Peer_Agent *agent, struct ATS_Address *address)
530 struct RIL_Address_Wrapped *cur;
532 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
533 if (cur->address_naked == address)
540 agent_action_is_possible (struct RIL_Peer_Agent *agent, int action)
546 case RIL_ACTION_NOTHING:
549 case RIL_ACTION_BW_IN_INC:
550 case RIL_ACTION_BW_IN_DBL:
551 if (agent->bw_in >= RIL_MAX_BW)
556 case RIL_ACTION_BW_IN_DEC:
557 case RIL_ACTION_BW_IN_HLV:
558 if (agent->bw_in <= 0)
563 case RIL_ACTION_BW_OUT_INC:
564 case RIL_ACTION_BW_OUT_DBL:
565 if (agent->bw_out >= RIL_MAX_BW)
570 case RIL_ACTION_BW_OUT_DEC:
571 case RIL_ACTION_BW_OUT_HLV:
572 if (agent->bw_out <= 0)
578 if ((action >= RIL_ACTION_TYPE_NUM) && (action < agent->n)) //switch address action
580 address_index = action - RIL_ACTION_TYPE_NUM;
582 GNUNET_assert(address_index >= 0);
584 address_index <= agent_address_get_index (agent, agent->addresses_tail->address_naked));
586 if ((agent_address_get_index(agent, agent->address_inuse) == address_index) ||
587 agent->address_inuse->active)
593 // error - action does not exist
594 GNUNET_assert(GNUNET_NO);
600 * Gets the action, with the maximal estimated Q-value (i.e. the one currently estimated to bring the
601 * most reward in the future)
603 * @param agent agent performing the calculation
604 * @param state the state from which to take the action
605 * @return the action promising most future reward
608 agent_get_action_max (struct RIL_Peer_Agent *agent, double *state)
611 int max_i = RIL_ACTION_INVALID;
613 double max_q = -DBL_MAX;
615 for (i = 0; i < agent->n; i++)
617 if (agent_action_is_possible(agent, i))
619 cur_q = agent_q (agent, state, i);
628 GNUNET_assert(RIL_ACTION_INVALID != max_i);
634 * Chooses a random action from the set of possible ones
636 * @param agent the agent performing the action
637 * @return the action index
640 agent_get_action_random (struct RIL_Peer_Agent *agent)
643 int is_possible[agent->n];
647 for (i = 0; i<agent->n; i++)
649 if (agent_action_is_possible(agent, i))
651 is_possible[i] = GNUNET_YES;
656 is_possible[i] = GNUNET_NO;
660 r = GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK, sum);
663 for (i = 0; i<agent->n; i++)
673 GNUNET_assert(GNUNET_NO);
674 return RIL_ACTION_INVALID;
679 * Updates the weights (i.e. coefficients) of the weight vector in matrix W for action a
681 * @param agent the agent performing the update
682 * @param reward the reward received for the last action
683 * @param s_next the new state, the last step got the agent into
684 * @param a_prime the new
687 agent_update (struct RIL_Peer_Agent *agent, double reward, double *s_next, int a_prime)
692 double **theta = agent->W;
694 delta = agent->envi->global_discount_integrated * reward; //reward
695 delta += agent->envi->global_discount_variable * agent_q (agent, s_next, a_prime); //discounted future value
696 delta -= agent_q (agent, agent->s_old, agent->a_old); //one step
698 // 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",
699 // agent->step_count,
700 // agent_q (agent, agent->s_old, agent->a_old),
701 // agent->envi->parameters.alpha,
703 // agent->envi->global_discount_variable,
704 // agent_q (agent, s_next, a_prime),
707 for (k = 0; k < agent->n; k++)
709 for (i = 0; i < agent->m; i++)
711 // LOG(GNUNET_ERROR_TYPE_INFO, "alpha = %f delta = %f e[%d] = %f\n",
712 // agent->envi->parameters.alpha,
716 theta[k][i] += agent->envi->parameters.alpha * delta * agent->E[k][i];
723 * Changes the eligibility trace vector e in various manners:
724 * #RIL_E_ACCUMULATE - adds @a feature to each component as in accumulating eligibility traces
725 * #RIL_E_REPLACE - resets each component to @a feature as in replacing traces
726 * #RIL_E_DECAY - multiplies e with discount factor and lambda as in the update rule
727 * #RIL_E_ZERO - sets e to 0 as in Watkin's Q-learning algorithm when exploring and when initializing
729 * @param agent the agent handle
730 * @param mod the kind of modification
731 * @param feature the feature vector
732 * @param action the action to take
735 agent_modify_eligibility (struct RIL_Peer_Agent *agent,
736 enum RIL_E_Modification mod,
743 for (i = 0; i < agent->m; i++)
747 case RIL_E_ACCUMULATE:
748 agent->E[action][i] += feature[i];
751 agent->E[action][i] = agent->E[action][i] > feature[i] ? agent->E[action][i] : feature[i];
754 for (k = 0; k < agent->n; k++)
756 agent->E[k][i] *= agent->envi->global_discount_variable * agent->envi->parameters.lambda;
760 for (k = 0; k < agent->n; k++)
770 * Informs the environment about the status of the solver
777 ril_inform (struct GAS_RIL_Handle *solver,
778 enum GAS_Solver_Operation op,
779 enum GAS_Solver_Status stat)
781 solver->env->info_cb (solver->env->cls,
788 * Calculates the maximum bandwidth an agent can assign in a network scope
792 static unsigned long long
793 ril_get_max_bw (struct RIL_Scope *net)
795 return GNUNET_MIN(2 * GNUNET_MAX(net->bw_in_available, net->bw_out_available), GNUNET_ATS_MaxBandwidth);
799 * Changes the active assignment suggestion of the handler and invokes the bw_changed callback to
800 * notify ATS of its new decision
802 * @param solver solver handle
803 * @param agent agent handle
804 * @param new_address the address which is to be used
805 * @param new_bw_in the new amount of inbound bandwidth set for this address
806 * @param new_bw_out the new amount of outbound bandwidth set for this address
807 * @param silent disables invocation of the bw_changed callback, if #GNUNET_YES
810 envi_set_active_suggestion (struct GAS_RIL_Handle *solver,
811 struct RIL_Peer_Agent *agent,
812 struct ATS_Address *new_address,
813 unsigned long long new_bw_in,
814 unsigned long long new_bw_out,
817 int notify = GNUNET_NO;
819 LOG(GNUNET_ERROR_TYPE_DEBUG,
820 " set_active_suggestion() for peer '%s'\n",
821 GNUNET_i2s (&agent->peer));
824 if (agent->address_inuse != new_address)
826 if (NULL != agent->address_inuse)
828 agent->address_inuse->active = GNUNET_NO;
829 agent->address_inuse->assigned_bw_in = 0;
830 agent->address_inuse->assigned_bw_out = 0;
832 if (NULL != new_address)
834 LOG(GNUNET_ERROR_TYPE_DEBUG, " set address active: %s\n", agent->is_active ? "yes" : "no");
835 new_address->active = agent->is_active;
836 new_address->assigned_bw_in = agent->bw_in;
837 new_address->assigned_bw_out = agent->bw_out;
839 notify |= GNUNET_YES;
845 if (new_address->active != agent->is_active)
847 new_address->active = agent->is_active;
848 notify |= GNUNET_YES;
852 if (agent->bw_in != new_bw_in)
854 agent->bw_in = new_bw_in;
855 new_address->assigned_bw_in = new_bw_in;
856 notify |= GNUNET_YES;
858 if (agent->bw_out != new_bw_out)
860 agent->bw_out = new_bw_out;
861 new_address->assigned_bw_out = new_bw_out;
862 notify |= GNUNET_YES;
866 if (notify && agent->is_active && (GNUNET_NO == silent))
870 LOG(GNUNET_ERROR_TYPE_DEBUG, " envi_set_active_suggestion() notify\n");
871 agent->suggestion_issue = GNUNET_YES;
872 agent->suggestion_address = new_address;
874 else if (agent->address_inuse)
876 /* disconnect case, no new address */
877 GNUNET_assert(0 == agent->address_inuse->assigned_bw_in);
878 GNUNET_assert(0 == agent->address_inuse->assigned_bw_out);
882 agent->suggestion_issue = GNUNET_YES;
883 agent->suggestion_address = agent->address_inuse;
886 agent->address_inuse = new_address;
891 * Allocates a state vector and fills it with the features present
892 * @param solver the solver handle
893 * @param agent the agent handle
894 * @return pointer to the state vector
897 envi_get_state (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
908 unsigned long long max_bw;
910 state = GNUNET_malloc (sizeof(double) * agent->m);
912 max_bw = ril_get_max_bw((struct RIL_Scope *) agent->address_inuse->solver_information);
914 y[0] = (double) agent->bw_out;
915 y[1] = (double) agent->bw_in;
917 m = agent_address_get_index (agent, agent->address_inuse) * (solver->parameters.rbf_divisor+1) * (solver->parameters.rbf_divisor+1);
918 for (i = 0; i <= solver->parameters.rbf_divisor; i++)
920 for (k = 0; k <= solver->parameters.rbf_divisor; k++)
922 x[0] = (double) i * (double) max_bw / (double) solver->parameters.rbf_divisor;
923 x[1] = (double) k * (double) max_bw / (double) solver->parameters.rbf_divisor;
926 sigma = (((double) max_bw / ((double) solver->parameters.rbf_divisor + 1)) * 0.5);
927 f = exp(-((d[0]*d[0] + d[1]*d[1]) / (2 * sigma * sigma)));
937 * Returns the utility value of the connection an agent manages
939 * @param agent the agent in question
940 * @return the utility value
943 agent_get_utility (struct RIL_Peer_Agent *agent)
945 const double *preferences;
950 preferences = agent->envi->env->get_preferences (agent->envi->env->cls,
953 delay_atsi = agent->address_inuse->norm_delay.norm;
954 delay_norm = RIL_UTILITY_DELAY_MAX*exp(-delay_atsi*0.00001);
956 pref_match = preferences[GNUNET_ATS_PREFERENCE_LATENCY] * delay_norm;
957 pref_match += preferences[GNUNET_ATS_PREFERENCE_BANDWIDTH] *
958 sqrt((double) (agent->bw_in/RIL_MIN_BW) * (double) (agent->bw_out/RIL_MIN_BW));
963 * Calculates the social welfare within a network scope according to what social
964 * welfare measure is set in the configuration.
966 * @param solver the solver handle
967 * @param scope the network scope in question
968 * @return the social welfare value
971 ril_network_get_social_welfare (struct GAS_RIL_Handle *solver, struct RIL_Scope *scope)
973 struct RIL_Peer_Agent *cur;
976 switch (solver->parameters.social_welfare)
978 case RIL_WELFARE_EGALITARIAN:
980 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
982 if (cur->is_active && cur->address_inuse && (cur->address_inuse->solver_information == scope))
984 result = GNUNET_MIN(result, agent_get_utility(cur));
989 case RIL_WELFARE_NASH:
991 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
993 if (cur->is_active && cur->address_inuse && (cur->address_inuse->solver_information == scope))
995 result *= pow(agent_get_utility(cur), 1.0 / (double) scope->active_agent_count);
1000 GNUNET_assert(GNUNET_NO);
1005 envi_get_penalty (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
1007 struct RIL_Scope *net;
1008 unsigned long long over_max;
1009 unsigned long long over_in = 0;
1010 unsigned long long over_out = 0;
1012 net = agent->address_inuse->solver_information;
1014 if (net->bw_in_utilized > net->bw_in_available)
1016 over_in = net->bw_in_utilized - net->bw_in_available;
1017 if (RIL_ACTION_BW_IN_INC == agent->a_old)
1019 /* increase quadratically */
1023 if (net->bw_out_utilized > net->bw_out_available)
1025 over_out = net->bw_out_utilized - net->bw_out_available;
1026 if (RIL_ACTION_BW_OUT_INC == agent->a_old)
1028 /* increase quadratically */
1029 over_out *= over_out;
1032 over_max = (over_in + over_out) / (RIL_MIN_BW * RIL_MIN_BW);
1034 return -1.0 * (double) over_max;
1038 * Gets the reward for the last performed step, which is calculated in equal
1039 * parts from the local (the peer specific) and the global (for all peers
1040 * identical) reward.
1042 * @param solver the solver handle
1043 * @param agent the agent handle
1044 * @return the reward
1047 envi_get_reward (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
1049 struct RIL_Scope *net;
1056 net = agent->address_inuse->solver_information;
1058 penalty = envi_get_penalty(solver, agent);
1059 objective = (agent_get_utility (agent) + net->social_welfare) / 2;
1060 delta = objective - agent->objective_old;
1061 agent->objective_old = objective;
1063 if (delta != 0 && penalty == 0)
1065 agent->nop_bonus = delta * RIL_NOP_DECAY;
1069 agent->nop_bonus *= RIL_NOP_DECAY;
1072 steady = (RIL_ACTION_NOTHING == agent->a_old) ? agent->nop_bonus : 0;
1074 reward = delta + steady;
1075 return reward + penalty;
1079 * Doubles the bandwidth for the active address
1081 * @param solver solver handle
1082 * @param agent agent handle
1083 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise the outbound bandwidth
1086 envi_action_bw_double (struct GAS_RIL_Handle *solver,
1087 struct RIL_Peer_Agent *agent,
1090 unsigned long long new_bw;
1091 unsigned long long max_bw;
1093 max_bw = ril_get_max_bw((struct RIL_Scope *) agent->address_inuse->solver_information);
1097 new_bw = agent->bw_in * 2;
1098 if (new_bw < agent->bw_in || new_bw > max_bw)
1100 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1101 agent->bw_out, GNUNET_NO);
1105 new_bw = agent->bw_out * 2;
1106 if (new_bw < agent->bw_out || new_bw > max_bw)
1108 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1114 * Cuts the bandwidth for the active address in half. The least amount of bandwidth suggested, is
1115 * the minimum bandwidth for a peer, in order to not invoke a disconnect.
1117 * @param solver solver handle
1118 * @param agent agent handle
1119 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1123 envi_action_bw_halven (struct GAS_RIL_Handle *solver,
1124 struct RIL_Peer_Agent *agent,
1127 unsigned long long new_bw;
1131 new_bw = agent->bw_in / 2;
1132 if (new_bw <= 0 || new_bw > agent->bw_in)
1134 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1139 new_bw = agent->bw_out / 2;
1140 if (new_bw <= 0 || new_bw > agent->bw_out)
1142 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1148 * Increases the bandwidth by 5 times the minimum bandwidth for the active address.
1150 * @param solver solver handle
1151 * @param agent agent handle
1152 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1156 envi_action_bw_inc (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1158 unsigned long long new_bw;
1159 unsigned long long max_bw;
1161 max_bw = ril_get_max_bw((struct RIL_Scope *) agent->address_inuse->solver_information);
1165 new_bw = agent->bw_in + (RIL_INC_DEC_STEP_SIZE * RIL_MIN_BW);
1166 if (new_bw < agent->bw_in || new_bw > max_bw)
1168 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw,
1169 agent->bw_out, GNUNET_NO);
1173 new_bw = agent->bw_out + (RIL_INC_DEC_STEP_SIZE * RIL_MIN_BW);
1174 if (new_bw < agent->bw_out || new_bw > max_bw)
1176 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in,
1182 * Decreases the bandwidth by 5 times the minimum bandwidth for the active address. The least amount
1183 * of bandwidth suggested, is the minimum bandwidth for a peer, in order to not invoke a disconnect.
1185 * @param solver solver handle
1186 * @param agent agent handle
1187 * @param direction_in if GNUNET_YES, change inbound bandwidth, otherwise change the outbound
1191 envi_action_bw_dec (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int direction_in)
1193 unsigned long long new_bw;
1197 new_bw = agent->bw_in - (RIL_INC_DEC_STEP_SIZE * RIL_MIN_BW);
1198 if (new_bw <= 0 || new_bw > agent->bw_in)
1200 envi_set_active_suggestion (solver, agent, agent->address_inuse, new_bw, agent->bw_out,
1205 new_bw = agent->bw_out - (RIL_INC_DEC_STEP_SIZE * RIL_MIN_BW);
1206 if (new_bw <= 0 || new_bw > agent->bw_out)
1208 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, new_bw,
1214 * Switches to the address given by its index
1216 * @param solver solver handle
1217 * @param agent agent handle
1218 * @param address_index index of the address as it is saved in the agent's list, starting with zero
1221 envi_action_address_switch (struct GAS_RIL_Handle *solver,
1222 struct RIL_Peer_Agent *agent,
1223 unsigned int address_index)
1225 struct RIL_Address_Wrapped *cur;
1228 //cur = agent_address_get_wrapped(agent, agent->address_inuse);
1230 for (cur = agent->addresses_head; NULL != cur; cur = cur->next)
1232 if (i == address_index)
1234 envi_set_active_suggestion (solver, agent, cur->address_naked, agent->bw_in, agent->bw_out,
1242 //no address with address_index exists, in this case this action should not be callable
1243 GNUNET_assert(GNUNET_NO);
1247 * Puts the action into effect by calling the according function
1249 * @param solver the solver handle
1250 * @param agent the action handle
1251 * @param action the action to perform by the solver
1254 envi_do_action (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int action)
1260 case RIL_ACTION_NOTHING:
1262 case RIL_ACTION_BW_IN_DBL:
1263 envi_action_bw_double (solver, agent, GNUNET_YES);
1265 case RIL_ACTION_BW_IN_HLV:
1266 envi_action_bw_halven (solver, agent, GNUNET_YES);
1268 case RIL_ACTION_BW_IN_INC:
1269 envi_action_bw_inc (solver, agent, GNUNET_YES);
1271 case RIL_ACTION_BW_IN_DEC:
1272 envi_action_bw_dec (solver, agent, GNUNET_YES);
1274 case RIL_ACTION_BW_OUT_DBL:
1275 envi_action_bw_double (solver, agent, GNUNET_NO);
1277 case RIL_ACTION_BW_OUT_HLV:
1278 envi_action_bw_halven (solver, agent, GNUNET_NO);
1280 case RIL_ACTION_BW_OUT_INC:
1281 envi_action_bw_inc (solver, agent, GNUNET_NO);
1283 case RIL_ACTION_BW_OUT_DEC:
1284 envi_action_bw_dec (solver, agent, GNUNET_NO);
1287 if ((action >= RIL_ACTION_TYPE_NUM) && (action < agent->n)) //switch address action
1289 address_index = action - RIL_ACTION_TYPE_NUM;
1291 GNUNET_assert(address_index >= 0);
1293 address_index <= agent_address_get_index (agent, agent->addresses_tail->address_naked));
1295 envi_action_address_switch (solver, agent, address_index);
1298 // error - action does not exist
1299 GNUNET_assert(GNUNET_NO);
1304 * Selects the next action using the e-greedy strategy. I.e. with a probability
1305 * of (1-e) the action with the maximum expected return will be chosen
1306 * (=> exploitation) and with probability (e) a random action will be chosen.
1307 * In case the Q-learning rule is set, the function also resets the eligibility
1308 * traces in the exploration case (after Watkin's Q-learning).
1310 * @param agent the agent selecting an action
1311 * @param state the current state-feature vector
1312 * @return the action index
1315 agent_select_egreedy (struct RIL_Peer_Agent *agent, double *state)
1318 double r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
1319 UINT32_MAX) / (double) UINT32_MAX;
1321 if (r < agent->envi->parameters.epsilon) //explore
1323 action = agent_get_action_random(agent);
1324 if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
1326 agent->eligibility_reset = GNUNET_YES;
1328 agent->envi->parameters.epsilon *= agent->envi->parameters.epsilon_decay;
1333 action = agent_get_action_max(agent, state);
1339 * Selects the next action with a probability corresponding to its value. The
1340 * probability is calculated using a Boltzmann distribution with a temperature
1341 * value. The higher the temperature, the more are the action selection
1342 * probabilities the same. With a temperature of 0, the selection is greedy,
1343 * i.e. always the action with the highest value is chosen.
1349 agent_select_softmax (struct RIL_Peer_Agent *agent, double *state)
1353 double eqt[agent->n];
1358 a_max = agent_get_action_max(agent, state);
1360 for (i=0; i<agent->n; i++)
1362 if (agent_action_is_possible(agent, i))
1364 eqt[i] = exp(agent_q(agent,state,i) / agent->envi->parameters.temperature);
1366 eqt[i] = isinf(eqt[i]) * UINT32_MAX;
1370 for (i=0; i<agent->n; i++)
1372 if (agent_action_is_possible(agent, i))
1381 r = (double) GNUNET_CRYPTO_random_u32 (GNUNET_CRYPTO_QUALITY_WEAK,
1382 UINT32_MAX) / (double) UINT32_MAX;
1384 for (i=0; i<agent->n; i++)
1390 if (RIL_ALGO_Q == agent->envi->parameters.algorithm)
1391 agent->eligibility_reset = GNUNET_YES;
1392 agent->envi->parameters.temperature *= agent->envi->parameters.temperature_decay;
1398 GNUNET_assert(GNUNET_NO);
1399 return RIL_ACTION_INVALID;
1403 * Select the next action of an agent either according to the e-greedy strategy
1404 * or the softmax strategy.
1406 * @param agent the agent in question
1407 * @param state the current state-feature vector
1408 * @return the action index
1411 agent_select_action (struct RIL_Peer_Agent *agent, double *state)
1413 if (agent->envi->parameters.select == RIL_SELECT_EGREEDY)
1415 return agent_select_egreedy(agent, state);
1419 return agent_select_softmax(agent, state);
1424 * Performs one step of the Markov Decision Process. Other than in the literature the step starts
1425 * after having done the last action a_old. It observes the new state s_next and the reward
1426 * received. Then the coefficient update is done according to the SARSA or Q-learning method. The
1427 * next action is put into effect.
1429 * @param agent the agent performing the step
1432 agent_step (struct RIL_Peer_Agent *agent)
1434 int a_next = RIL_ACTION_INVALID;
1439 LOG(GNUNET_ERROR_TYPE_DEBUG, " agent_step() Peer '%s', algorithm %s\n",
1440 GNUNET_i2s (&agent->peer),
1441 agent->envi->parameters.algorithm ? "Q" : "SARSA");
1443 s_next = envi_get_state (agent->envi, agent);
1444 reward = envi_get_reward (agent->envi, agent);
1446 if (agent->eligibility_reset)
1448 agent_modify_eligibility(agent, RIL_E_ZERO, NULL, -1);
1449 agent->eligibility_reset = GNUNET_NO;
1453 agent_modify_eligibility (agent, RIL_E_DECAY, NULL, -1);
1455 if (RIL_ACTION_INVALID != agent->a_old)
1457 agent_modify_eligibility (agent, agent->envi->parameters.eligibility_trace_mode, agent->s_old, agent->a_old);
1460 switch (agent->envi->parameters.algorithm)
1462 case RIL_ALGO_SARSA:
1463 a_next = agent_select_action (agent, s_next);
1464 if (RIL_ACTION_INVALID != agent->a_old)
1466 //updates weights with selected action (on-policy), if not first step
1467 agent_update (agent, reward, s_next, a_next);
1472 a_max = agent_get_action_max (agent, s_next);
1473 if (RIL_ACTION_INVALID != agent->a_old)
1475 //updates weights with best action, disregarding actually selected action (off-policy), if not first step
1476 agent_update (agent, reward, s_next, a_max);
1478 a_next = agent_select_action (agent, s_next);
1482 GNUNET_assert(RIL_ACTION_INVALID != a_next);
1484 LOG (GNUNET_ERROR_TYPE_DEBUG, "step() Step# %llu R: %f IN %llu OUT %llu A: %d\n",
1491 envi_do_action (agent->envi, agent, a_next);
1493 GNUNET_free(agent->s_old);
1494 agent->s_old = s_next;
1495 agent->a_old = a_next;
1497 agent->step_count += 1;
1501 * Prototype of the ril_step() procedure
1503 * @param solver the solver handle
1506 ril_step (struct GAS_RIL_Handle *solver);
1510 * Task for the scheduler, which performs one step and lets the solver know that
1511 * no further step is scheduled.
1513 * @param cls the solver handle
1516 ril_step_scheduler_task (void *cls)
1518 struct GAS_RIL_Handle *solver = cls;
1520 solver->step_next_task_id = NULL;
1525 * Determines how much of the available bandwidth is assigned. If more is
1526 * assigned than available it returns 1. The function is used to determine the
1527 * step size of the adaptive stepping.
1529 * @param solver the solver handle
1533 ril_get_used_resource_ratio (struct GAS_RIL_Handle *solver)
1536 struct RIL_Scope net;
1537 unsigned long long sum_assigned = 0;
1538 unsigned long long sum_available = 0;
1541 for (i = 0; i < solver->networks_count; i++)
1543 net = solver->network_entries[i];
1544 if (net.bw_in_assigned > 0) //only consider scopes where an address is actually active
1546 sum_assigned += net.bw_in_utilized;
1547 sum_assigned += net.bw_out_utilized;
1548 sum_available += net.bw_in_available;
1549 sum_available += net.bw_out_available;
1552 if (sum_available > 0)
1554 ratio = ((double) sum_assigned) / ((double) sum_available);
1561 return ratio > 1 ? 1 : ratio; //overutilization is possible, cap at 1
1565 * Lookup network struct by type
1567 * @param s the solver handle
1568 * @param type the network type
1569 * @return the network struct
1571 static struct RIL_Scope *
1572 ril_get_network (struct GAS_RIL_Handle *s, uint32_t type)
1576 for (i = 0; i < s->networks_count; i++)
1578 if (s->network_entries[i].type == type)
1580 return &s->network_entries[i];
1587 * Determines whether more connections are allocated in a network scope, than
1588 * they would theoretically fit. This is used as a heuristic to determine,
1589 * whether a new connection can be allocated or not.
1591 * @param solver the solver handle
1592 * @param network the network scope in question
1593 * @return GNUNET_YES if there are theoretically enough resources left
1596 ril_network_is_not_full (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
1598 struct RIL_Scope *net;
1599 struct RIL_Peer_Agent *agent;
1600 unsigned long long address_count = 0;
1602 for (agent = solver->agents_head; NULL != agent; agent = agent->next)
1604 if (agent->address_inuse && agent->is_active)
1606 net = agent->address_inuse->solver_information;
1607 if (net->type == network)
1614 net = ril_get_network (solver, network);
1615 return (net->bw_in_available > RIL_MIN_BW * address_count) && (net->bw_out_available > RIL_MIN_BW * address_count);
1619 * Unblocks an agent for which a connection request is there, that could not
1620 * be satisfied. Iterates over the addresses of the agent, if one of its
1621 * addresses can now be allocated in its scope the agent is unblocked,
1622 * otherwise it remains unchanged.
1624 * @param solver the solver handle
1625 * @param agent the agent in question
1629 ril_try_unblock_agent (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent, int silent)
1631 struct RIL_Address_Wrapped *addr_wrap;
1632 struct RIL_Scope *net;
1633 unsigned long long start_in;
1634 unsigned long long start_out;
1636 for (addr_wrap = agent->addresses_head; NULL != addr_wrap; addr_wrap = addr_wrap->next)
1638 net = addr_wrap->address_naked->solver_information;
1639 if (ril_network_is_not_full(solver, net->type))
1641 if (NULL == agent->address_inuse)
1643 start_in = net->bw_in_available < net->bw_in_utilized ? (net->bw_in_available - net->bw_in_utilized) / 2 : RIL_MIN_BW;
1644 start_out = net->bw_out_available < net->bw_out_utilized ? (net->bw_out_available - net->bw_out_utilized) / 2 : RIL_MIN_BW;
1645 envi_set_active_suggestion (solver, agent, addr_wrap->address_naked, start_in, start_out, silent);
1650 agent->address_inuse = NULL;
1654 * Determines how much the reward needs to be discounted depending on the amount
1655 * of time, which has passed since the last time-step.
1657 * @param solver the solver handle
1660 ril_calculate_discount (struct GAS_RIL_Handle *solver)
1662 struct GNUNET_TIME_Absolute time_now;
1663 struct GNUNET_TIME_Relative time_delta;
1666 // MDP case only for debugging purposes
1667 if (solver->simulate)
1669 solver->global_discount_variable = solver->parameters.gamma;
1670 solver->global_discount_integrated = 1;
1676 //calculate tau, i.e. how many real valued time units have passed, one time unit is one minimum time step
1677 time_now = GNUNET_TIME_absolute_get ();
1678 time_delta = GNUNET_TIME_absolute_get_difference (solver->step_time_last, time_now);
1679 solver->step_time_last = time_now;
1680 tau = (double) time_delta.rel_value_us
1681 / (double) solver->parameters.step_time_min.rel_value_us;
1683 //calculate reward discounts (once per step for all agents)
1684 solver->global_discount_variable = pow (M_E, ((-1.0) * ((double) solver->parameters.beta) * tau));
1685 solver->global_discount_integrated = (1.0 - solver->global_discount_variable)
1686 / (double) solver->parameters.beta;
1690 * Count the number of active agents/connections in a network scope
1692 * @param solver the solver handle
1693 * @param scope the network scope in question
1694 * @return the number of allocated connections
1697 ril_network_count_active_agents (struct GAS_RIL_Handle *solver, struct RIL_Scope *scope)
1700 struct RIL_Peer_Agent *cur_agent;
1702 for (cur_agent = solver->agents_head; NULL != cur_agent; cur_agent = cur_agent->next)
1704 if (cur_agent->is_active && cur_agent->address_inuse && (cur_agent->address_inuse->solver_information == scope))
1713 * Calculates how much bandwidth is assigned in sum in a network scope, either
1714 * in the inbound or in the outbound direction.
1716 * @param solver the solver handle
1717 * @param type the type of the network scope in question
1718 * @param direction_in GNUNET_YES if the inbound direction should be summed up,
1719 * otherwise the outbound direction will be summed up
1720 * @return the sum of the assigned bandwidths
1722 static unsigned long long
1723 ril_network_get_assigned (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
1725 struct RIL_Peer_Agent *cur;
1726 struct RIL_Scope *net;
1727 unsigned long long sum = 0;
1729 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1731 if (cur->is_active && cur->address_inuse)
1733 net = cur->address_inuse->solver_information;
1734 if (net->type == type)
1748 * Calculates how much bandwidth is actually utilized in sum in a network scope,
1749 * either in the inbound or in the outbound direction.
1751 * @param solver the solver handle
1752 * @param type the type of the network scope in question
1753 * @param direction_in GNUNET_YES if the inbound direction should be summed up,
1754 * otherwise the outbound direction will be summed up
1755 * @return the sum of the utilized bandwidths (in bytes/second)
1757 static unsigned long long
1758 ril_network_get_utilized (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type type, int direction_in)
1760 struct RIL_Peer_Agent *cur;
1761 struct RIL_Scope *net;
1762 unsigned long long sum = 0;
1764 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1766 if (cur->is_active && cur->address_inuse)
1768 net = cur->address_inuse->solver_information;
1769 if (net->type == type)
1772 sum += cur->address_inuse->norm_utilization_in.norm;
1774 sum += cur->address_inuse->norm_utilization_out.norm;
1783 * Retrieves the state of the network scope, so that its attributes are up-to-
1786 * @param solver the solver handle
1789 ril_networks_update_state (struct GAS_RIL_Handle *solver)
1792 struct RIL_Scope *net;
1794 for (c = 0; c < solver->networks_count; c++)
1796 net = &solver->network_entries[c];
1797 net->bw_in_assigned = ril_network_get_assigned(solver, net->type, GNUNET_YES);
1798 net->bw_in_utilized = ril_network_get_utilized(solver, net->type, GNUNET_YES);
1799 net->bw_out_assigned = ril_network_get_assigned(solver, net->type, GNUNET_NO);
1800 net->bw_out_utilized = ril_network_get_utilized(solver, net->type, GNUNET_NO);
1801 net->active_agent_count = ril_network_count_active_agents(solver, net);
1802 net->social_welfare = ril_network_get_social_welfare(solver, net);
1807 * Schedules the next global step in an adaptive way. The more resources are
1808 * left, the earlier the next step is scheduled. This serves the reactivity of
1809 * the solver to changed inputs.
1811 * @param solver the solver handle
1814 ril_step_schedule_next (struct GAS_RIL_Handle *solver)
1820 struct GNUNET_TIME_Relative time_next;
1822 used_ratio = ril_get_used_resource_ratio (solver);
1825 solver->parameters.step_time_min.rel_value_us
1826 <= solver->parameters.step_time_max.rel_value_us);
1828 factor = (double) GNUNET_TIME_relative_subtract (solver->parameters.step_time_max,
1829 solver->parameters.step_time_min).rel_value_us;
1830 offset = (double) solver->parameters.step_time_min.rel_value_us;
1831 y = factor * pow (used_ratio, RIL_INTERVAL_EXPONENT) + offset;
1833 GNUNET_assert(y <= (double) solver->parameters.step_time_max.rel_value_us);
1834 GNUNET_assert(y >= (double) solver->parameters.step_time_min.rel_value_us);
1836 time_next = GNUNET_TIME_relative_saturating_multiply (GNUNET_TIME_UNIT_MICROSECONDS, (unsigned long long) y);
1838 // LOG (GNUNET_ERROR_TYPE_INFO, "ratio: %f, factor: %f, offset: %f, y: %f\n",
1844 if (solver->simulate)
1846 time_next = GNUNET_TIME_UNIT_ZERO;
1849 if ((NULL == solver->step_next_task_id) && (GNUNET_NO == solver->done))
1851 solver->step_next_task_id = GNUNET_SCHEDULER_add_delayed (time_next, &ril_step_scheduler_task,
1857 * Triggers one step per agent
1862 ril_step (struct GAS_RIL_Handle *solver)
1864 struct RIL_Peer_Agent *cur;
1866 if (GNUNET_YES == solver->bulk_lock)
1868 solver->bulk_changes++;
1872 ril_inform (solver, GAS_OP_SOLVE_START, GAS_STAT_SUCCESS);
1874 LOG(GNUNET_ERROR_TYPE_DEBUG, " RIL step number %d\n", solver->step_count);
1876 if (0 == solver->step_count)
1878 solver->step_time_last = GNUNET_TIME_absolute_get ();
1881 ril_calculate_discount (solver);
1882 ril_networks_update_state (solver);
1884 //trigger one step per active, unblocked agent
1885 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1889 if (NULL == cur->address_inuse)
1891 ril_try_unblock_agent(solver, cur, GNUNET_NO);
1893 if (cur->address_inuse)
1900 ril_networks_update_state (solver);
1902 solver->step_count++;
1903 ril_step_schedule_next (solver);
1905 ril_inform (solver, GAS_OP_SOLVE_STOP, GAS_STAT_SUCCESS);
1907 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_START, GAS_STAT_SUCCESS);
1908 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
1910 if (cur->suggestion_issue) {
1911 solver->env->bandwidth_changed_cb (solver->env->cls,
1912 cur->suggestion_address);
1913 cur->suggestion_issue = GNUNET_NO;
1916 ril_inform (solver, GAS_OP_SOLVE_UPDATE_NOTIFICATION_STOP, GAS_STAT_SUCCESS);
1920 * Initializes the matrix W of parameter vectors theta with small random numbers.
1922 * @param agent The respective agent
1925 agent_w_init (struct RIL_Peer_Agent *agent)
1930 for (i = 0; i < agent->n; i++)
1932 for (k = 0; k < agent->m; k++)
1934 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));
1940 * Initialize an agent without addresses and its knowledge base
1942 * @param s ril solver
1943 * @param peer the one in question
1944 * @return handle to the new agent
1946 static struct RIL_Peer_Agent *
1947 agent_init (void *s, const struct GNUNET_PeerIdentity *peer)
1950 struct GAS_RIL_Handle * solver = s;
1951 struct RIL_Peer_Agent * agent = GNUNET_new (struct RIL_Peer_Agent);
1953 agent->envi = solver;
1954 agent->peer = *peer;
1955 agent->step_count = 0;
1956 agent->is_active = GNUNET_NO;
1957 agent->bw_in = RIL_MIN_BW;
1958 agent->bw_out = RIL_MIN_BW;
1959 agent->suggestion_issue = GNUNET_NO;
1960 agent->n = RIL_ACTION_TYPE_NUM;
1962 agent->W = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
1963 agent->E = (double **) GNUNET_malloc (sizeof (double *) * agent->n);
1964 for (i = 0; i < agent->n; i++)
1966 agent->W[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1967 agent->E[i] = (double *) GNUNET_malloc (sizeof (double) * agent->m);
1969 agent_w_init(agent);
1970 agent->eligibility_reset = GNUNET_NO;
1971 agent->a_old = RIL_ACTION_INVALID;
1972 agent->s_old = GNUNET_malloc (sizeof (double) * agent->m);
1973 agent->address_inuse = NULL;
1974 agent->objective_old = 0;
1975 agent->nop_bonus = 0;
1983 * @param solver the solver handle
1984 * @param agent the agent to retire
1987 agent_die (struct GAS_RIL_Handle *solver, struct RIL_Peer_Agent *agent)
1991 for (i = 0; i < agent->n; i++)
1993 GNUNET_free_non_null(agent->W[i]);
1994 GNUNET_free_non_null(agent->E[i]);
1996 GNUNET_free_non_null(agent->W);
1997 GNUNET_free_non_null(agent->E);
1998 GNUNET_free_non_null(agent->s_old);
2003 * Returns the agent for a peer
2005 * @param solver the solver handle
2006 * @param peer the identity of the peer
2007 * @param create whether or not to create an agent, if none is allocated yet
2010 static struct RIL_Peer_Agent *
2011 ril_get_agent (struct GAS_RIL_Handle *solver, const struct GNUNET_PeerIdentity *peer, int create)
2013 struct RIL_Peer_Agent *cur;
2015 for (cur = solver->agents_head; NULL != cur; cur = cur->next)
2017 if (0 == memcmp (peer, &cur->peer, sizeof(struct GNUNET_PeerIdentity)))
2025 cur = agent_init (solver, peer);
2026 GNUNET_CONTAINER_DLL_insert_tail(solver->agents_head, solver->agents_tail, cur);
2033 * Determine whether at least the minimum bandwidth is set for the network. Otherwise the network is
2034 * considered inactive and not used. Addresses in an inactive network are ignored.
2036 * @param solver solver handle
2037 * @param network the network type
2038 * @return whether or not the network is considered active
2041 ril_network_is_active (struct GAS_RIL_Handle *solver, enum GNUNET_ATS_Network_Type network)
2043 struct RIL_Scope *net;
2045 net = ril_get_network (solver, network);
2046 return net->bw_out_available >= RIL_MIN_BW;
2050 * Cuts a slice out of a vector of elements. This is used to decrease the size of the matrix storing
2051 * the reward function approximation. It copies the memory, which is not cut, to the new vector,
2052 * frees the memory of the old vector, and redirects the pointer to the new one.
2054 * @param old pointer to the pointer to the first element of the vector
2055 * @param element_size byte size of the vector elements
2056 * @param hole_start the first element to cut out
2057 * @param hole_length the number of elements to cut out
2058 * @param old_length the length of the old vector
2061 ril_cut_from_vector (void **old,
2062 size_t element_size,
2063 unsigned int hole_start,
2064 unsigned int hole_length,
2065 unsigned int old_length)
2068 char *oldptr = (char *) *old;
2070 unsigned int bytes_before;
2071 unsigned int bytes_hole;
2072 unsigned int bytes_after;
2074 GNUNET_assert(old_length >= hole_length);
2075 GNUNET_assert(old_length >= (hole_start + hole_length));
2077 size = element_size * (old_length - hole_length);
2079 bytes_before = element_size * hole_start;
2080 bytes_hole = element_size * hole_length;
2081 bytes_after = element_size * (old_length - hole_start - hole_length);
2089 tmpptr = GNUNET_malloc (size);
2090 GNUNET_memcpy (tmpptr, oldptr, bytes_before);
2091 GNUNET_memcpy (tmpptr + bytes_before, oldptr + (bytes_before + bytes_hole), bytes_after);
2097 *old = (void *) tmpptr;
2101 * Solver API functions
2102 * ---------------------------
2106 * Change relative preference for quality in solver
2108 * @param solver the solver handle
2109 * @param peer the peer to change the preference for
2110 * @param kind the kind to change the preference
2111 * @param pref_rel the normalized preference value for this kind over all clients
2114 GAS_ril_address_change_preference (void *solver,
2115 const struct GNUNET_PeerIdentity *peer,
2116 enum GNUNET_ATS_PreferenceKind kind,
2119 LOG(GNUNET_ERROR_TYPE_DEBUG,
2120 "API_address_change_preference() Preference '%s' for peer '%s' changed to %.2f \n",
2121 GNUNET_ATS_print_preference_type (kind), GNUNET_i2s (peer), pref_rel);
2123 struct GAS_RIL_Handle *s = solver;
2125 s->parameters.temperature = s->parameters.temperature_init;
2126 s->parameters.epsilon = s->parameters.epsilon_init;
2132 * Add a new address for a peer to the solver
2134 * The address is already contained in the addresses hashmap!
2136 * @param solver the solver Handle
2137 * @param address the address to add
2138 * @param network network type of this address
2141 GAS_ril_address_add (void *solver,
2142 struct ATS_Address *address,
2145 struct GAS_RIL_Handle *s = solver;
2146 struct RIL_Peer_Agent *agent;
2147 struct RIL_Address_Wrapped *address_wrapped;
2148 struct RIL_Scope *net;
2156 LOG (GNUNET_ERROR_TYPE_DEBUG,
2157 "API_address_add()\n");
2159 net = ril_get_network (s, network);
2160 address->solver_information = net;
2162 if (!ril_network_is_active (s, network))
2164 LOG(GNUNET_ERROR_TYPE_DEBUG,
2165 "API_address_add() Did not add %s address %s for peer '%s', network does not have enough bandwidth\n",
2166 address->plugin, address->addr, GNUNET_i2s (&address->peer));
2170 s->parameters.temperature = s->parameters.temperature_init;
2171 s->parameters.epsilon = s->parameters.epsilon_init;
2173 agent = ril_get_agent (s, &address->peer, GNUNET_YES);
2176 address_wrapped = GNUNET_new (struct RIL_Address_Wrapped);
2177 address_wrapped->address_naked = address;
2178 GNUNET_CONTAINER_DLL_insert_tail(agent->addresses_head, agent->addresses_tail, address_wrapped);
2180 //increase size of W
2181 m_new = agent->m + ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1));
2183 n_new = agent->n + 1;
2186 GNUNET_array_grow(agent->W, agent->n, n_new);
2188 GNUNET_array_grow(agent->E, agent->n, n_new);
2189 for (i = 0; i < n_new; i++)
2194 GNUNET_array_grow(agent->W[i], agent->m, m_new);
2196 GNUNET_array_grow(agent->E[i], agent->m, m_new);
2201 GNUNET_array_grow(agent->W[i], zero, m_new);
2203 GNUNET_array_grow(agent->E[i], zero, m_new);
2207 //increase size of old state vector
2209 GNUNET_array_grow(agent->s_old, agent->m, m_new);
2211 ril_try_unblock_agent(s, agent, GNUNET_NO);
2215 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_address_add() Added %s %s address %s for peer '%s'\n",
2216 address->active ? "active" : "inactive", address->plugin, address->addr,
2217 GNUNET_i2s (&address->peer));
2221 * Delete an address in the solver
2223 * The address is not contained in the address hashmap anymore!
2225 * @param solver the solver handle
2226 * @param address the address to remove
2229 GAS_ril_address_delete (void *solver,
2230 struct ATS_Address *address)
2232 struct GAS_RIL_Handle *s = solver;
2233 struct RIL_Peer_Agent *agent;
2234 struct RIL_Address_Wrapped *address_wrapped;
2239 struct RIL_Scope *net;
2241 LOG (GNUNET_ERROR_TYPE_DEBUG,
2242 "API_address_delete() Delete %s %s address %s for peer '%s'\n",
2243 address->active ? "active" : "inactive",
2246 GNUNET_i2s (&address->peer));
2248 agent = ril_get_agent (s, &address->peer, GNUNET_NO);
2251 net = address->solver_information;
2252 GNUNET_assert(! ril_network_is_active (s, net->type));
2253 LOG (GNUNET_ERROR_TYPE_DEBUG,
2254 "No agent allocated for peer yet, since address was in inactive network\n");
2258 s->parameters.temperature = s->parameters.temperature_init;
2259 s->parameters.epsilon = s->parameters.epsilon_init;
2261 address_index = agent_address_get_index (agent, address);
2262 address_wrapped = agent_address_get_wrapped (agent, address);
2264 if (NULL == address_wrapped)
2266 net = address->solver_information;
2267 LOG (GNUNET_ERROR_TYPE_DEBUG,
2268 "Address not considered by agent, address was in inactive network\n");
2271 GNUNET_CONTAINER_DLL_remove (agent->addresses_head,
2272 agent->addresses_tail,
2274 GNUNET_free (address_wrapped);
2277 m_new = agent->m - ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1));
2278 n_new = agent->n - 1;
2280 for (i = 0; i < agent->n; i++)
2282 ril_cut_from_vector ((void **) &agent->W[i], sizeof(double),
2283 address_index * ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)),
2284 ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)), agent->m);
2285 ril_cut_from_vector ((void **) &agent->E[i], sizeof(double),
2286 address_index * ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)),
2287 ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)), agent->m);
2289 GNUNET_free_non_null(agent->W[RIL_ACTION_TYPE_NUM + address_index]);
2290 GNUNET_free_non_null(agent->E[RIL_ACTION_TYPE_NUM + address_index]);
2291 ril_cut_from_vector ((void **) &agent->W, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
2293 ril_cut_from_vector ((void **) &agent->E, sizeof(double *), RIL_ACTION_TYPE_NUM + address_index,
2295 //correct last action
2296 if (agent->a_old > (RIL_ACTION_TYPE_NUM + address_index))
2300 else if (agent->a_old == (RIL_ACTION_TYPE_NUM + address_index))
2302 agent->a_old = RIL_ACTION_INVALID;
2304 //decrease old state vector
2305 ril_cut_from_vector ((void **) &agent->s_old, sizeof(double),
2306 address_index * ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)),
2307 ((s->parameters.rbf_divisor+1) * (s->parameters.rbf_divisor+1)), agent->m);
2311 if (agent->address_inuse == address)
2313 if (NULL != agent->addresses_head) //if peer has an address left, use it
2315 GNUNET_log (GNUNET_ERROR_TYPE_DEBUG,
2316 "Active address died, suggesting alternative!\n");
2317 envi_set_active_suggestion (s,
2319 agent->addresses_head->address_naked,
2326 GNUNET_log (GNUNET_ERROR_TYPE_DEBUG,
2327 "Active address died, suggesting disconnect!\n");
2328 envi_set_active_suggestion (s, agent, NULL, 0, 0, GNUNET_NO);
2332 if (agent->suggestion_address == address)
2334 agent->suggestion_issue = GNUNET_NO;
2335 agent->suggestion_address = NULL;
2337 GNUNET_assert (agent->address_inuse != address);
2342 * Update the properties of an address in the solver
2344 * @param solver solver handle
2345 * @param address the address
2348 GAS_ril_address_property_changed (void *solver,
2349 struct ATS_Address *address)
2351 struct GAS_RIL_Handle *s = solver;
2353 LOG(GNUNET_ERROR_TYPE_DEBUG,
2354 "Properties for peer '%s' address changed\n",
2355 GNUNET_i2s (&address->peer));
2356 s->parameters.temperature = s->parameters.temperature_init;
2357 s->parameters.epsilon = s->parameters.epsilon_init;
2363 * Give feedback about the current assignment
2365 * @param solver the solver handle
2366 * @param application the application
2367 * @param peer the peer to change the preference for
2368 * @param scope the time interval for this feedback: [now - scope .. now]
2369 * @param kind the kind to change the preference
2370 * @param score the score
2373 GAS_ril_address_preference_feedback (void *solver,
2374 struct GNUNET_SERVICE_Client *application,
2375 const struct GNUNET_PeerIdentity *peer,
2376 const struct GNUNET_TIME_Relative scope,
2377 enum GNUNET_ATS_PreferenceKind kind,
2380 LOG (GNUNET_ERROR_TYPE_DEBUG,
2381 "API_address_preference_feedback() Peer '%s' got a feedback of %+.3f from application %s for "
2382 "preference %s for %d seconds\n",
2385 GNUNET_ATS_print_preference_type (kind),
2386 scope.rel_value_us / 1000000);
2391 * Start a bulk operation
2393 * @param solver the solver
2396 GAS_ril_bulk_start (void *solver)
2398 struct GAS_RIL_Handle *s = solver;
2400 LOG (GNUNET_ERROR_TYPE_DEBUG,
2401 "API_bulk_start() lock: %d\n", s->bulk_lock+1);
2408 * Bulk operation done
2410 * @param solver the solver handle
2413 GAS_ril_bulk_stop (void *solver)
2415 struct GAS_RIL_Handle *s = solver;
2417 LOG(GNUNET_ERROR_TYPE_DEBUG,
2418 "API_bulk_stop() lock: %d\n",
2421 if (s->bulk_lock < 1)
2428 if (0 < s->bulk_changes)
2431 s->bulk_changes = 0;
2437 * Tell solver to notify ATS if the address to use changes for a specific
2438 * peer using the bandwidth changed callback
2440 * The solver must only notify about changes for peers with pending address
2443 * @param solver the solver handle
2444 * @param peer the identity of the peer
2447 GAS_ril_get_preferred_address (void *solver,
2448 const struct GNUNET_PeerIdentity *peer)
2450 struct GAS_RIL_Handle *s = solver;
2451 struct RIL_Peer_Agent *agent;
2453 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_get_preferred_address()\n");
2455 agent = ril_get_agent (s, peer, GNUNET_YES);
2457 agent->is_active = GNUNET_YES;
2458 envi_set_active_suggestion (solver, agent, agent->address_inuse, agent->bw_in, agent->bw_out, GNUNET_YES);
2460 ril_try_unblock_agent(solver, agent, GNUNET_YES);
2462 if (agent->address_inuse)
2464 LOG(GNUNET_ERROR_TYPE_DEBUG,
2465 "API_get_preferred_address() Activated agent for peer '%s' with %s address %s\n",
2466 GNUNET_i2s (peer), agent->address_inuse->plugin, agent->address_inuse->addr);
2470 LOG(GNUNET_ERROR_TYPE_DEBUG,
2471 "API_get_preferred_address() Activated agent for peer '%s', but no address available\n",
2473 s->parameters.temperature = s->parameters.temperature_init;
2474 s->parameters.epsilon = s->parameters.epsilon_init;
2476 if (NULL != agent->address_inuse)
2477 s->env->bandwidth_changed_cb (s->env->cls,
2478 agent->address_inuse);
2483 * Tell solver stop notifying ATS about changes for this peers
2485 * The solver must only notify about changes for peers with pending address
2488 * @param solver the solver handle
2489 * @param peer the peer
2492 GAS_ril_stop_get_preferred_address (void *solver,
2493 const struct GNUNET_PeerIdentity *peer)
2495 struct GAS_RIL_Handle *s = solver;
2496 struct RIL_Peer_Agent *agent;
2498 LOG (GNUNET_ERROR_TYPE_DEBUG,
2499 "API_stop_get_preferred_address()");
2501 agent = ril_get_agent (s, peer, GNUNET_NO);
2508 if (GNUNET_NO == agent->is_active)
2514 s->parameters.temperature = s->parameters.temperature_init;
2515 s->parameters.epsilon = s->parameters.epsilon_init;
2517 agent->is_active = GNUNET_NO;
2519 envi_set_active_suggestion (s, agent, agent->address_inuse, agent->bw_in, agent->bw_out,
2524 LOG(GNUNET_ERROR_TYPE_DEBUG,
2525 "API_stop_get_preferred_address() Paused agent for peer '%s'\n",
2531 * Entry point for the plugin
2533 * @param cls pointer to the 'struct GNUNET_ATS_PluginEnvironment'
2536 libgnunet_plugin_ats_ril_init (void *cls)
2538 static struct GNUNET_ATS_SolverFunctions sf;
2539 struct GNUNET_ATS_PluginEnvironment *env = cls;
2540 struct GAS_RIL_Handle *solver = GNUNET_new (struct GAS_RIL_Handle);
2541 struct RIL_Scope * cur;
2546 LOG (GNUNET_ERROR_TYPE_DEBUG,
2547 "API_init() Initializing RIL solver\n");
2549 GNUNET_assert (NULL != env);
2550 GNUNET_assert (NULL != env->cfg);
2551 GNUNET_assert (NULL != env->stats);
2552 GNUNET_assert (NULL != env->bandwidth_changed_cb);
2553 GNUNET_assert (NULL != env->get_preferences);
2555 if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number(env->cfg, "ats", "RIL_RBF_DIVISOR", &solver->parameters.rbf_divisor))
2557 solver->parameters.rbf_divisor = RIL_DEFAULT_RBF_DIVISOR;
2561 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MIN",
2562 &solver->parameters.step_time_min))
2564 solver->parameters.step_time_min = RIL_DEFAULT_STEP_TIME_MIN;
2568 != GNUNET_CONFIGURATION_get_value_time (env->cfg, "ats", "RIL_STEP_TIME_MAX",
2569 &solver->parameters.step_time_max))
2571 solver->parameters.step_time_max = RIL_DEFAULT_STEP_TIME_MAX;
2574 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_ALGORITHM", &string))
2576 GNUNET_STRINGS_utf8_toupper (string, string);
2577 if (0 == strcmp (string, "SARSA"))
2579 solver->parameters.algorithm = RIL_ALGO_SARSA;
2581 if (0 == strcmp (string, "Q-LEARNING"))
2583 solver->parameters.algorithm = RIL_ALGO_Q;
2586 GNUNET_free (string);
2590 solver->parameters.algorithm = RIL_DEFAULT_ALGORITHM;
2593 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_SELECT", &string))
2595 solver->parameters.select = !strcmp (string, "EGREEDY") ? RIL_SELECT_EGREEDY : RIL_SELECT_SOFTMAX;
2596 GNUNET_free (string);
2600 solver->parameters.select = RIL_DEFAULT_SELECT;
2604 solver->parameters.beta = RIL_DEFAULT_DISCOUNT_BETA;
2605 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2606 "RIL_DISCOUNT_BETA", &f_tmp))
2610 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2611 "RIL_DISCOUNT_BETA", f_tmp);
2615 solver->parameters.beta = f_tmp;
2616 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2617 "RIL_DISCOUNT_BETA", f_tmp);
2621 solver->parameters.gamma = RIL_DEFAULT_DISCOUNT_GAMMA;
2622 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2623 "RIL_DISCOUNT_GAMMA", &f_tmp))
2625 if ((f_tmp < 0.0) || (f_tmp > 1.0))
2627 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2628 "RIL_DISCOUNT_GAMMA", f_tmp);
2632 solver->parameters.gamma = f_tmp;
2633 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2634 "RIL_DISCOUNT_GAMMA", f_tmp);
2638 solver->parameters.alpha = RIL_DEFAULT_GRADIENT_STEP_SIZE;
2639 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2640 "RIL_GRADIENT_STEP_SIZE", &f_tmp))
2642 if ((f_tmp < 0.0) || (f_tmp > 1.0))
2644 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2645 "RIL_GRADIENT_STEP_SIZE", f_tmp);
2649 solver->parameters.alpha = f_tmp;
2650 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2651 "RIL_GRADIENT_STEP_SIZE", f_tmp);
2655 solver->parameters.lambda = RIL_DEFAULT_TRACE_DECAY;
2656 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2657 "RIL_TRACE_DECAY", &f_tmp))
2659 if ((f_tmp < 0.0) || (f_tmp > 1.0))
2661 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2662 "RIL_TRACE_DECAY", f_tmp);
2666 solver->parameters.lambda = f_tmp;
2667 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2668 "RIL_TRACE_DECAY", f_tmp);
2672 solver->parameters.epsilon_init = RIL_DEFAULT_EXPLORE_RATIO;
2673 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2674 "RIL_EXPLORE_RATIO", &f_tmp))
2676 if ((f_tmp < 0.0) || (f_tmp > 1.0))
2678 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2679 "RIL_EXPLORE_RATIO", f_tmp);
2683 solver->parameters.epsilon_init = f_tmp;
2684 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2685 "RIL_EXPLORE_RATIO", f_tmp);
2689 solver->parameters.epsilon_decay = RIL_DEFAULT_EXPLORE_DECAY;
2690 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2691 "RIL_EXPLORE_DECAY", &f_tmp))
2693 if ((f_tmp < 0.0) || (f_tmp > 1.0))
2695 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2696 "RIL_EXPLORE_DECAY", f_tmp);
2700 solver->parameters.epsilon_decay = f_tmp;
2701 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2702 "RIL_EXPLORE_DECAY", f_tmp);
2706 solver->parameters.temperature_init = RIL_DEFAULT_TEMPERATURE;
2707 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2708 "RIL_TEMPERATURE", &f_tmp))
2712 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2713 "RIL_TEMPERATURE", f_tmp);
2717 solver->parameters.temperature_init = f_tmp;
2718 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2719 "RIL_TEMPERATURE", f_tmp);
2723 solver->parameters.temperature_decay = RIL_DEFAULT_TEMPERATURE_DECAY;
2724 if (GNUNET_SYSERR != GNUNET_CONFIGURATION_get_value_float (env->cfg, "ats",
2725 "RIL_TEMPERATURE_DECAY", &f_tmp))
2727 if ((f_tmp <= 0.0) || solver->parameters.temperature_decay > 1)
2729 LOG (GNUNET_ERROR_TYPE_ERROR, _("Invalid %s configuration %f \n"),
2730 "RIL_TEMPERATURE_DECAY", f_tmp);
2734 solver->parameters.temperature_decay = f_tmp;
2735 LOG (GNUNET_ERROR_TYPE_INFO, "Using %s of %.3f\n",
2736 "RIL_TEMPERATURE_DECAY", f_tmp);
2740 if (GNUNET_OK != GNUNET_CONFIGURATION_get_value_number (env->cfg, "ats", "RIL_SIMULATE", &solver->simulate))
2742 solver->simulate = GNUNET_NO;
2745 if (GNUNET_YES == GNUNET_CONFIGURATION_get_value_yesno(env->cfg, "ats", "RIL_REPLACE_TRACES"))
2747 solver->parameters.eligibility_trace_mode = RIL_E_REPLACE;
2751 solver->parameters.eligibility_trace_mode = RIL_E_ACCUMULATE;
2754 if (GNUNET_OK == GNUNET_CONFIGURATION_get_value_string (env->cfg, "ats", "RIL_SOCIAL_WELFARE", &string))
2756 solver->parameters.social_welfare = !strcmp (string, "NASH") ? RIL_WELFARE_NASH : RIL_WELFARE_EGALITARIAN;
2757 GNUNET_free (string);
2761 solver->parameters.social_welfare = RIL_DEFAULT_WELFARE;
2766 sf.s_add = &GAS_ril_address_add;
2767 sf.s_address_update_property = &GAS_ril_address_property_changed;
2768 sf.s_get = &GAS_ril_get_preferred_address;
2769 sf.s_get_stop = &GAS_ril_stop_get_preferred_address;
2770 sf.s_pref = &GAS_ril_address_change_preference;
2771 sf.s_feedback = &GAS_ril_address_preference_feedback;
2772 sf.s_del = &GAS_ril_address_delete;
2773 sf.s_bulk_start = &GAS_ril_bulk_start;
2774 sf.s_bulk_stop = &GAS_ril_bulk_stop;
2776 solver->networks_count = env->network_count;
2777 solver->network_entries = GNUNET_malloc (env->network_count * sizeof (struct RIL_Scope));
2778 solver->step_count = 0;
2779 solver->done = GNUNET_NO;
2781 for (c = 0; c < env->network_count; c++)
2783 cur = &solver->network_entries[c];
2785 cur->bw_in_available = env->in_quota[c];
2786 cur->bw_out_available = env->out_quota[c];
2787 LOG (GNUNET_ERROR_TYPE_DEBUG,
2788 "init() Quotas for %s network: IN %llu - OUT %llu\n",
2789 GNUNET_ATS_print_network_type(cur->type),
2790 cur->bw_in_available/1024,
2791 cur->bw_out_available/1024);
2794 LOG(GNUNET_ERROR_TYPE_DEBUG, "init() Parameters:\n");
2795 LOG(GNUNET_ERROR_TYPE_DEBUG, "init() Algorithm = %s, alpha = %f, beta = %f, lambda = %f\n",
2796 solver->parameters.algorithm ? "Q" : "SARSA",
2797 solver->parameters.alpha,
2798 solver->parameters.beta,
2799 solver->parameters.lambda);
2800 LOG(GNUNET_ERROR_TYPE_DEBUG, "init() exploration_ratio = %f, temperature = %f, ActionSelection = %s\n",
2801 solver->parameters.epsilon,
2802 solver->parameters.temperature,
2803 solver->parameters.select ? "EGREEDY" : "SOFTMAX");
2804 LOG(GNUNET_ERROR_TYPE_DEBUG, "init() RBF_DIVISOR = %llu\n",
2805 solver->parameters.rbf_divisor);
2812 * Exit point for the plugin
2814 * @param cls the solver handle
2817 libgnunet_plugin_ats_ril_done (void *cls)
2819 struct GNUNET_ATS_SolverFunctions *sf = cls;
2820 struct GAS_RIL_Handle *s = sf->cls;
2821 struct RIL_Peer_Agent *cur_agent;
2822 struct RIL_Peer_Agent *next_agent;
2824 LOG(GNUNET_ERROR_TYPE_DEBUG, "API_done() Shutting down RIL solver\n");
2826 s->done = GNUNET_YES;
2828 cur_agent = s->agents_head;
2829 while (NULL != cur_agent)
2831 next_agent = cur_agent->next;
2832 GNUNET_CONTAINER_DLL_remove(s->agents_head, s->agents_tail, cur_agent);
2833 agent_die (s, cur_agent);
2834 cur_agent = next_agent;
2837 if (NULL != s->step_next_task_id)
2839 GNUNET_SCHEDULER_cancel (s->step_next_task_id);
2841 GNUNET_free(s->network_entries);
2848 /* end of plugin_ats_ril.c */