Artificial Minds with Consciousness and Common sense Aspects

The research work presented in this article investigates and explains the conceptual mechanisms of consciousness and common-sense thinking of animates. These mechanisms are computationally simulated on artificial agents as strategic rules to analyze and compare the performance of agents in critical and dynamic environments. Awareness and attention to specific parameters that affect the performance of agents specify the consciousness level in agents. Common sense is a set of beliefs that are accepted to be true among a group of agents that are engaged in a common purpose, with or without self-experience. The common sense agents are a kind of conscious agents that are given with few common sense assumptions. The so-created environment has attackers with dependency on agents in the survival-food chain. These attackers create a threat mental state in agents that can affect their conscious and common sense behaviors. The agents are built with a multi-layer cognitive architecture COCOCA (Consciousness and Common sense Cognitive Architecture) with five columns and six layers of cognitive processing of each precept of an agent. The conscious agents self-learn strategies for threat management and energy level maintenance. Experimentation conducted in this research work demonstrates animate-level intelligence in their problem-solving capabilities, decision making and reasoning in critical situations.


Theory of Conscious Agents
Mostofthehumanmentalprocessesareunconsciousthoughhumansareconsidered ashighlyconsciousagents (Bargh&Morsella,2008).Theconsciousagentsarethe entitiesthatexhibitintelligentbehaviorwithpropertiessuchasautonomy,reactiveness, and pro-activeness or being rational. According to Donald D Hoffman (2014), the mathematical definition of a conscious agent involves three mental processes such asperception,decisionmaking,andaction.Anagentbeinginaconsciousstatecan also have subjective experiences, wishes, beliefs, desires, and complex thoughts (Block,1995;Shoemaker,1996). It should be able to understand a relatively complex sequence of actions at an abstract level and respond to such situations (Franklin, 2009). A minimum prerequisite for conscious agents is social interactionwithitspeersintheenvironment.

Theories of Consciousness
AccordingtotheBDImodelproposedbyBratman(1988),thepracticalreasoning processofhumanshastwosteps:(a)considerallthedesiresofanagentand(b)select the most desirable one by mapping it to its current belief set. In this deliberated step, the agent pursues and adopts an intention to achieve a desire. The intentions arepersistentinnatureandrecurtilltheyareachieved.Iftheintentionchosenfails repeatedlytoachievethedesiredstate,theagentcandropthisandupdateitsbelief set. Hence, intentions are the prime reason for an agent to change its future belief set.Ineachintentionalstateanagentconsidersoradoptsoptionsthatareconsistent withthatintention.Inprinciple,intentionsjustifythepossibilityofachievingagoal state in the current state. The second step in practical reasoning process involves generatingaplanofactionsbasedongoals,beliefs,andactionsofagentsbyusing means-ends reasoning. A running agent adopts varying plans that are triggered by externalorinternalevents.Thisplaninvolvesasequenceofactionsthatareselected basedontheavailablesetofbeliefs.
The attackers always look for a nearby agent as a prey to satisfy their food requirement.Theattackersensurethattheydonotattackagentswheningroupand agentswithcommonsenseusethisknowledgetoescapethethreat.Asafe-zoneis alsocreatedinanenvironmentwheretheattackerdoesnotenterandthisisknownto thecommonsenseagentstoo.
In the deliberative layer the precepts are mapped on to agents' Belief-Desire-Intention (BDI) to trigger motivated actions. The attention selector processes in theconsciousnesslayerevaluatethemotivatedactionsinthedeliberativelayerand update the belief set frequently. The self-reflective layer monitors every conscious actiontriggeredandtheireffectontheagent'sinternalstateandexternalworld.This formsafeedbackfor convertingsome of thebeliefs as common sense beliefs. The parametersthatmayaffecttheimmediategoalorcantriggerfearasanemotioninan agentarepushedintoglobalworkspaceorworkingmemorytogetconsciouscontrol. The emotions and motivations play a major role in generating consciousness and commonsenseinanagent'sbehavior.Thisleadsanagenttomanageitsmotivations andgoalsbyselectingappropriatestrategies.Thestrategiescanbeeitherconscious orcommonsensethatisbasedonthemeta-reasoninglogicadoptedbyanagent.The default meta-reasoning is: if in normal scenarios, common sense triggers and if in fear, consciousness improves. These strategies selected are constructed into action setandsenttoanactiongenerator,whichinturnchangestheexternalenvironment (seeFigure3).
Theperceivedinputsareprocessedinthislayertoformassociationwithcurrent the set of beliefs. The cross-product of the belief set and desire set gives a set of intentionsthatarepossibleinthecurrentstate.Theconsciousprocessinhigherlayer weighstheintentionsetandchoosestheonewithhigherweightasadeliberatedaction. The belief set of an agent is defined with the facts about the environment and the self.Thissetinitiallycontainsthefactsoftheexternalworldsuchastheavailability ofrawfruits,dryfruits,juicyfruits,obstacles,andborderofthearena.Thesetalso includesitsinternalparameterssuchasitsenergylevel,state,direction,name,and color.Thedesiresetisasetofactionsthatanagentcanexecutebasedontheactuator set at its disposal. There is a subset of beliefs that are defined as common sense beliefs.Commonsenseagentsmayhavethesamesetofdesiresbutuseadifferent setofbeliefsforreasoning.

design of Conscious Agents
Theconsciouslayerisdesignedbyusingtheaxiomatic (Aleksander,I.&Dunmall, 2003;Aleksander, I., 2007) theory by simulating cognitive abilities that make an agent conscious. The conscious agents are built by using proactive attitudes like beliefs,desires,intentions,andemotionsthatformthebasisformotivatedactions, whichareconsciousbydefault.Theagentsaregivenacollectionofstrategiesthat suitsdifferentenvironmentalconditions.Theconsciousagents,basedontheirgoals, choosetherequiredparameters,whichareinfocalattention,fromtheperceptionlist. Thechangesintheenvironmentmakeanagenttriggerdifferentactionsbychoosing different strategies. Each conscious strategy depicts different cognitive abilities to demonstrate consciousness levels in the agents. The conscious agents use internal affect value of objects in the external world and BDI set to choose the strategies. Thesestrategiesworkonapartialorderplanningtoaccomplishthegoalsassigned. TheinternalstructureoftheconsciousagentisshowninFigure6.
Theagentsintheconsciouslayerarebuiltbyusingbehaviorsinthedeliberative layer and in-turn in the reactive and reflexive layers. The state-transition diagram for conscious agent is as shown in Figure 7. The agent of this layer can be threatconscious,energy-conscious,orboth.Initially,theagentdemonstratesexplorebehavior  Step 1 Agent with Simple-Reflexive-Explore {Update-Energy} Step 2 Adopt default Conscious-Strategy { Update-Energy} Step 3 Monitor internal parameters { EMax Thr <= Energy <= EMin Thr } Step 4 If Energy <= EMax Thr go to IDLE state Step 5 Go to step 1 Agent behavior in the CONSCIOUS state: Step 1 If Energy <= EMin Thr Step 2 Initialize strategic-planning Step 3 Evaluate Pre-conditions {current energy-level, current position (Grid-location), current precepts} Step 4 Change Strategy {Choose next strategy available in list} Step 5 Monitor Energy if EMax Thr >= Energy >= EMin Thr i.
Initiate learn-fruit ii.
Update knowledge Else Go to step 4 Agent behavior in the IDLE state: Step 1 Agent in same position for next 5 cycles Step 2 ,a) is the cumulative reward gained in previous strategy adopted (+ve value if energy increased otherwise -ve) R(s, a') is the reward gained by applying current strategy P best is the previous best cumulative reward gained for any strategy adopted by an agent Best Avg is the feasible incremental reward expected after adopting a strategy Step 1: Initialize Q(s,a) =0 α = 0.1 ϒ = 0.9 Step2: Calculate Q(s',a') + = α * (R(s, a') + ϒ * (P best -Q(s,a)) Step3: if Q(s',a') > Best Avg Save strategy with precondition Best Avg =Q(s',a') Else if Q(s',a') < Bes Avg Unlearn strategy for current preconditions Step 4: s =s' and Q(s,a) =Q(s',a') Step 5: go to step 2
Thecommonsenseagentsadoptknownstrategiesfirstandthenlearntooptimize the behavior. The initial strategies are regularly updated by agents through their experiences.Thecommonsensestrategiesareusedbyagentswhentheyareaware thattheycanalwaysescapefromtheattacker.

Common Sense Agent's Behavior in the CONSCIOUS State for Threat Management
Step 1 If Threat-level = HIGH Step 2 Initialize strategic-planning Step 3 Evaluate Pre-conditions {Current Threat-level, current position (Gridlocation), current precepts} Step 4 Choose strategy from common sense by mapping pre-conditions Step 5 Monitor Threat-level if Threat -level == MEDIUM Continue with the same strategy Else if Threat-level == LOW Update current strategy as best for current pre-conditions Else if Threat-level == HIGH Delete the strategy from list Go to step 5

ReSULT ANALySIS oF CoCoCA AGeNTS IN SIMULATed ANIMATe TeSTBed
Theresultsarecapturedfromtheagent'sbehaviorinhandlingtheirinnerstatesin critical conditions. All agent types are evaluated for energy-level maintenance and escaping rates as performance metrics. Figure 9 is a graph that shows the energy-levelmaintenanceinFSM,energy-conscious,andconscious2agentsinasingle-agent environmentandintheabsenceofanattacker.Theseagentsareseparatelycompared forenergylevelsastheyconsciouslymonitorenergylevelsandusethesamebelief set.TheFSMagentistheleast-consciousagentanddoesnotmaintainenergylevels evenwhensufficientfruitsareavailableintheenvironmentasitalwaysgoeswith defaultstrategy. Theenergy-consciousagentmaintainsitsenergylevelonthreshold,anddoesnot consumerawfruitunlessitsenergyleveldropsbelowthethreshold.Thereinforcement learning in conscious agents enables energy conscious agents to achieve this by dynamicallychangingthestrategiesthataremoreprobabletoachievethegoalstate.
In the presence of an attacker, the agents are evaluated for escape-count. As energy-consciousagentsarenotthreat-conscious,theycanbekilledbyanattacker,but duringitssurvivalitsenergymaintenanceisbettercomparedtothethreat-conscious andFSMagents.

CoNCLUSIoN
Theresearchworkhighlightstheideathatmanyaspectsofconsciousnessandcommon sensethinkingcanbesimulatedonagents.Theresearchexperimenthasprogressively achievedtheresultsrequiredtojustifytheoutcomes.Theconceptualmechanismsof consciousness and common sense have been computationally represented by using cognitivearchitecturecalledCOCOCAwithsix-layers-five-columns.Experimentation isconductedbyusinganimatetestbedwithsimulatedagents.Theagentsaremade to survive in an environment with attacker coexistence. This makes an agent to consciouslyfocusonthethreatlevelineachmovetomakeadecisionforthenextmove.